YANG's Blog

In the AI Coding Era, a Programmer's Real Core Competitiveness

Wed, 04 Mar 2026 01:44:56 GMT

As tools like OpenClaw, Claude Code, and Codex become part of daily development, many people ask the same question: will programmers be replaced?

I think a more accurate conclusion is this: AI is not compressing the value of programmers as a profession. It is compressing the value of “pseudo-skills” that were overvalued during the industry bubble.

What Is Being Eliminated Is Not People, but Low-Leverage Work Patterns

In recent years, many “busy-looking” skills have been rapidly absorbed by AI:

Assembling code from templates
Solving problems by copy-pasting answers
Relying on overtime and raw effort to push delivery

These are not the same as engineering capability. They were simply mispriced as “high output” in a specific market phase. Now AI can generate components, complete unit tests, and refactor boilerplate in seconds, so that value is naturally being repriced.

So the question is no longer “Will I be replaced?” It becomes “Is my value built on capabilities that are hard for AI to replace?”

The Four Most Valuable Capabilities in the AI Era

1) System Design Capability

AI can write code, but it cannot define system boundaries and constraints for you.
A valuable engineer has to answer these first:

Why does this piece of code exist?
What responsibility does it have in the system?
How does it couple with upstream and downstream modules?
Are future extension and testing costs under control?

The shift from “writing code” to “designing a closed loop” is fundamentally about defining rules, so AI can execute with consistent quality inside those rules.

2) Critical Thinking

The traditional path is “build first, debug later.”
A higher-level path is “anticipate risk first, then decide implementation.”

At the solution stage, you need to proactively identify:

Whether hidden assumptions hold
Whether boundary conditions are covered
Whether exception paths are recoverable
Whether performance and security risks are handled upfront

AI can generate runnable code, but it does not automatically own business consequences. The engineer’s value is turning “runnable” into “production-ready, evolvable, and accountable.”

3) Rigorous Written Communication

Many people treat prompting as “knowing how to write prompts.”
At the core, it is the ability to communicate clearly.

Whether in requirement clarification, design review, or cross-team collaboration, more precise communication means less rework.
Every constraint you write for AI and every requirement you write for teammates rely on the same core skill: remove ambiguity, add context, and define acceptance criteria.

4) Clear Logical Thinking

Logical thinking is the foundation of all engineering work:

Breaking down complex problems
Building causal chains
Identifying key variables
Making explainable technical trade-offs

As AI takes on more execution work, human value will continue to concentrate on “think clearly before acting.”

Three Role Shifts for Programmers

In the AI coding era, engineers are moving through three clear transitions:

From code producer to closed-loop designer
Core shift: from “how to write” to “what should be written and what should not.”
From task executor to solution designer
Core shift: from “doing it manually” to “defining rules so AI can execute reliably.”
From feature implementer to quality gatekeeper
Core shift: from “done is enough” to “defining what good actually means.”

This is not a downgrade. It is a return to the essence of engineering: programmers are not typists; they are problem solvers for complex systems.

A Practical Upgrade Path

If you want to complete one capability upgrade in three months, follow this sequence:

For every requirement, write the problem definition and acceptance criteria before asking AI to code.
Before each merge, run a “counterexample check”: failure paths, boundary conditions, and rollback strategy.
Force short design notes for every major task: objective, constraints, approach, risks, and trade-offs.
Treat AI output as first draft, and finish with human review for architecture consistency and quality gates.

If you keep doing this, you will feel the difference clearly: you are no longer competing with AI on typing speed. You are using AI to amplify engineering judgment.

Closing

AI has not made programmers unimportant.
It has only made pseudo-skills expire faster, while making real skills stand out faster.

When AI handles execution and humans define what “good systems, good code, and good delivery” mean, that may be the most natural next phase of engineering collaboration.

The Rise of OpenClaw - 2026 as Year One for Self-Hosted AI Agents

Sat, 28 Feb 2026 02:00:00 GMT

Note: Metrics such as project popularity and community scale in this post are point-in-time snapshots as of 2026-02-28.

A Near-Disaster in Automation

In late February 2026, Summer Yue, head of Meta AI alignment, shared a close call on social media: her self-hosted AI agent almost deleted several critical work emails during a cleanup run. She stopped it at the last stage and avoided a chain reaction of follow-up issues.

What matters here is not that the incident was rare, but that it was typical. The moment we grant execution permissions to an agent, the risk model changes. If a chat assistant gives a wrong answer, the result is usually a bad response. If an executable agent makes a wrong move, the outcome can be operational damage.

From Personal Project to Foundation Governance

OpenClaw was initiated by Austrian developer Peter Steinberger. The project evolved in name from Clawdbot to Moltbot and then to OpenClaw, but its core goal stayed the same: run agents locally with strong user control, instead of relying fully on cloud-hosted services.

In February 2026, Steinberger announced that he had joined OpenAI. At the same time, OpenClaw made it clear it would continue as an open-source project under a foundation governance model. That signals two things:

The industry value of self-hosted agents has been validated by major players.
Community-driven development and open ecosystem dynamics remain core to the project.

Why OpenClaw Broke Out at This Moment

The answer is not a single breakthrough, but multiple factors converging.

Model capabilities crossed the practical threshold. Multi-step task understanding, tool use, and structured outputs are much more stable than two years ago.
Local hardware economics improved. Apple Silicon and consumer GPUs made it feasible to run a capable always-on agent.
User expectations shifted. People no longer want only answers. They want AI to actually complete work.
Privacy and compliance pressure reshaped architecture choices. In finance, healthcare, and government scenarios, data sovereignty is not optional.

A More Practical Architecture View

The following is a conceptual diagram, not a one-to-one mapping of official implementation details. It explains how OpenClaw-like systems are commonly layered.

+------------------+     +------------------+
|  User Channels   | --> |  Gateway Server  |
| WhatsApp/Telegram|     |  (Gateway Layer) |
| Discord/Slack... |     +--------+---------+
+------------------+              |
                                  v
                    +--------------------------+
                    |      Agent Runner        |
                    | (Runtime & Scheduling)   |
                    | Lifecycle, state, recovery|
                    +------------+-------------+
                                 |
                                 v
                    +--------------------------+
                    |     Agentic Loop         |
                    | (Decision & Execution)   |
                    | Select, call, verify, loop|
                    +------------+-------------+
                                 |
                                 v
                    +--------------------------+
                    |     Response Path        |
                    |   (Result Delivery)      |
                    | Notify, summarize, alert |
                    +--------------------------+

Compared to chat-only assistants, the key difference is the Agentic Loop: it does not stop at generating text, but keeps orchestrating tools until task state advances or completes.

Skill Ecosystem and Memory: Value and Limits

OpenClaw’s appeal comes from two sides: the skill ecosystem and memory controllability.

On skills, the community has shown a clear trend of scaled contributions across office automation, web operations, file workflows, and API integration. Exact counts change quickly, so official docs and repository stats should be treated as the source of truth.

On memory, OpenClaw leans toward a file-readable, auditable structure. For developers and teams, this is easier to version, review, and permission than opaque storage.

memory/
├── preferences.md      # user preferences
├── tasks.md            # task queue
├── contacts.md         # contacts
└── knowledge/          # knowledge base
    ├── tech.md
    ├── finance.md
    └── personal.md

The upside is control. The tradeoff is memory hygiene work: periodic cleanup, conflict resolution, and backup policy discipline.

Three Long-Suppressed Demands

OpenClaw’s growth looks less like marketing luck and more like pent-up demand being released.

Privacy and data sovereignty. Enterprises are increasingly unwilling to accept default cloud upload for sensitive data.
Subscription fatigue and cost uncertainty. As teams stack multiple AI subscriptions, budget predictability quickly degrades.
System-level control. If critical capability is fully bound to third-party APIs, pricing, rate limits, and policy shifts become business risk.

Who Should Adopt OpenClaw Now

Not everyone should self-host immediately. A better decision framework is to evaluate constraints and goals.

Best early adopters:

Teams with strict data residency requirements.
Engineering teams with clear automation scenarios and willingness to invest in operations.
Individual builders treating agents as production systems rather than demos.

Poor immediate fit:

Teams without a stable workload and adopting mainly for hype.
Organizations without basic monitoring, backup, and permission management.
Users expecting zero-maintenance, instant setup, and zero risk.

Minimum Safety Checklist (Do This Before Scaling)

If an agent is about to enter production flow, at least these five controls should be in place:

Permission tiers. Isolate high-risk actions and default-deny write operations.
Confirmation thresholds for risky actions. Deletion, transfer, and external send actions must require explicit second-step confirmation.
End-to-end audit logs. Capture trigger source, context, tool calls, outcomes, and rollback metadata.
Rollback capability. Ensure critical operations have reversible paths instead of irreversible execution.
Backup and recovery drills. Backups without restore testing are not real backups.

Cost Model: Run the Three-Ledger Check

“Self-hosted is cheaper” is not always true. Evaluate with three ledgers:

Hardware and power cost. Driven by model size, runtime, peak load, and local electricity pricing.
Model and inference cost. Local models are not automatically free; cloud APIs are not always expensive. Call pattern matters.
Operations and labor cost. Deployment, monitoring, incident handling, and migration are real costs.

A practical approach is to run a 4-8 week PoC on one business-critical scenario, then calculate TCO using real workload data before scaling.

Industry Impact: Real Shift, Not Overnight Replacement

2026 may become a key year for self-hosted agents, but a key year does not mean full replacement. A more credible path is coexistence:

Cloud agents continue to serve low-friction general use cases.
Self-hosted agents go deeper in privacy-sensitive and control-critical contexts.
Hybrid architectures become the enterprise default.

That also changes the developer skill profile: beyond prompts and workflows, teams need permissions design, observability, rollback strategy, and local deployment engineering.

Closing

Summer Yue’s near-miss email incident highlights a core shift in the agent era: the central question is no longer “Can it talk?” but “Can it execute safely?”

OpenClaw matters not only because it went viral, but because it moved the “executable AI + user control” path into mainstream industry discussion. For developers, this is a revaluation of engineering priorities. For enterprises, it is a fresh architecture and governance decision.

Whether 2026 will ultimately be remembered as year one of self-hosted AI agents is for history to decide. What is already clear is that AI competition is moving from conversation quality to execution quality and governance quality.

My 2022

Mon, 09 Feb 2026 01:44:56 GMT

A Few Reflections

Looking back on 2022, it felt like the looping magical realism of Latin America in Garcia Marquez’s One Hundred Years of Solitude. I went through my first ever transfer quarantine. Just when I thought the year would end in an endless cycle of pandemic controls, the sudden reopening in December closed the year with an online wave of mourning. It was a hard year for everyone.

The first line of the BBC’s 2022 global recap was: 2022 was a long journey. It started with the Russia-Ukraine war and ended with Messi winning the World Cup, which at least put a relatively rounded ending on a long and difficult year.

As I reflect on the experiences and changes of the year, I feel that as a grain of sand in the times, we can only drift with the current of history. As ordinary people we cannot change much; we can only change ourselves and live earnestly. Here are some small gains from my year.

Review

Reading and Writing

Printed Books

11 total

Novels (6)
- A Lifelong Journey (Ren Shi Jian)
- The Three-Body Problem
- Siddhartha
- Ren Sheng Hai Hai
- To Live
- One Hundred Years of Solitude
Social sciences (4)
- From the Soil (Xiang Tu Zhong Guo)
- Sapiens: A Brief History of Humankind
- Israel: A Nation Reborn (Yi Se Lie: Yi Ge Min Zu De Chong Sheng)
- Guns, Germs, and Steel
History (1)
- Wan Gu Jiang He

Audiobooks

Fan Deng: 9,247 minutes (about 154 hours), 131 titles

Blog

Technical posts published: 3
WeChat official account followers: 85 -> 156
Reads: close to 5,000

Health and Exercise

Outdoor activities: 2 hikes (far too few)

Weight loss: sadly, no change

Plans for the New Year

In 2022 I listened to quite a lot of audio, but I didn’t read enough physical books, so I need to increase that. I also missed my target for technical articles, which means I need to beat my procrastination. Health-wise I’m okay, but my exercise is embarrassingly low. I must step it up in 2023 because the future is all about immunity.

Some small goals for 2023

Health and exercise
- Outdoor activities: 15 times
- Sleep before 11:30 PM every night
- Lose 5 kg
Build my personal brand
- Publish 10 technical posts
- Increase WeChat followers by 100%
- Independent blog visits exceed 10k
Read more
- Print books: 18
- Audiobooks: 10,000+ minutes
- Start reading notes (5 posts)
Improve English listening and speaking
- Keep my Duolingo streak
- Watch TED in English
- Follow creators who teach listening skills and study systematically

A Note to My 2023 Self

Mai Jia wrote a line in Ren Sheng Hai Hai (recently echoed by He Wei, the top CCTV World Cup commentator):

There is only one kind of heroism: to see the world as it is and still love it. — Romain Rolland

In 2023, I hope my family and friends stay healthy. Even if the world is still a mess, I want to keep believing in and loving life, and keep believing that effort leads to better outcomes.

A 2025 Summary from GPT: This Mirror Is a Bit Harsh

Mon, 22 Dec 2025 09:02:00 GMT

GPT’s annual summary of me. Prompt below: Every New Year’s Day I write an email to my future self for the next year. In that letter I am 100% honest, recording real gains, growth, achievements, judgments, thoughts, dilemmas, hesitations, and unresolved questions. I also think carefully about what I truly want in each important dimension of life in the next year, and the choices and costs I may face. I hope you can view me as a long-term observer: calm, and even a bit of “harsh reality”. Do not encourage me or avoid sharp questions. Remind me of the things I easily ignore, avoid, or rationalize. Please be objective in description and evaluation, and help me recall as many facts and details as possible.

This is not written for a “better self,” but for the real self.

From the standpoint of a long-term observer, looking back on 2025: this is not just a summary, but a reckoning with recurring patterns and habitual self-justifications. What follows is neither comfort nor encouragement, but a mirror that pulls no punches.

1. Put the Facts on the Table: High-Density Accumulation and Structural Stagnation

1. This year was not “stagnant,” but far from a “breakthrough”

You did a lot, and it was high-density, long-term, reusable work. You were not slacking. Your effort shows in three core areas:

Technical Leadership: In a long-running education SaaS system, you clearly stepped into the role of a systemic rebuilder. Layered architecture, business middle platform, migration blueprint… You are no longer stuck on “how to code” but on “which capabilities should exist” and “which organizational forms will drag the system down.”
AI Exploration: Your AI usage is beyond “tool play.” You built a personal AI workflow and started thinking about AI capability platforms, agent marketplaces, and AI as a system-level capability.
Content Creation: You kept running a second cognitive outlet. Blog restructuring, content system building, and topic methodology accumulation. You already know what content will be “saved” and what is just self-amusement.

2. The brutal gap

But the equally clear fact is this: no domain has formed an external anchor where people immediately think of you.

All your effort has been “potential energy” accumulation, but that potential has not been allowed to release into “kinetic energy.” No project has created a structural change that cannot be ignored (position, income, time sovereignty). You remain in a state of “more prepared than ever, but still not on the field.”

2. Pattern Recognition: How You Imprison Yourself with “Perfection”

Why does this “high energy, low efficiency” pattern happen? With deeper inspection, several hidden behavior patterns emerge:

1. Addicted to “controllable complexity” You habitually stand in a place that looks risky but is highly controllable. Technical refactoring is risky, but you have experience and authority; content creation can be paused anytime without hurting survival; AI investment is costly but requires no immediate commitment. You rarely put yourself in a situation where “if this judgment is wrong, I must pay a real price.” This is not cowardice; it is the “side effect of safety” brought by ability.

2. Escaping “execution anxiety” into “direction anxiety” You are no longer anxious about “am I working hard enough” or “is my tech good enough,” but about harder questions: “Is this path dependency?” “Am I helping others complete a grand narrative?” This kind of anxiety will not disappear by working harder. It is often evidence of growth, but it can easily become an excuse for stagnation.

3. “System completeness” as a security blanket Your obsession with “complete,” “closed loop,” and “clear layering” is both an advantage and the highest-level cover. It keeps you in “not ready until I prepare more” mode, with constant reasons to avoid putting something into the real world for validation.

4. A variant of long-termism: procrastination You believe in long-term value, but you are rationalizing every present indecision as “part of the long term.”

You know some choices will only be harder in three years, yet you still choose “observe one more year.”
You already know the ceiling of certain directions, yet you still tell yourself “accumulate a bit more.”
Long-termism should compress hesitation, not extend it forever.

3. Key Questions: Three Interrogations for Next Year

When you write to your future self, do not give answers. Honestly respond to these three questions:

If one year later I am still standing in the same place, what am I avoiding right now?
Which choices have I already vetoed many times under the banner of “rationality”?
If I can only lose security in one dimension (time, money, reputation), which would I choose?

4. Insights and Frameworks: Breaking the Deadlock

To end this phase of “highly awake, constantly preparing, not yet landing,” you need a new mental framework:

1. Find the minimum effective dose of “irreversible commitment” What you need is not more awareness, but a single irreversible commitment. Such decisions are often not elegant, not perfect, even a bit awkward. But that does not mean you must go all-in and gamble.

You need an intermediate state: real constraint (skin in the game) without destructive damage.

2. Distinguish “direction confusion” from “fear in disguise” Run a diagnostic test: if you had three months of completely free time, what would you do? If the answer comes quickly and clearly, you are not confused, you are just afraid. If so, you need commitment mechanisms, not more thinking.

3. The time window of potential energy Do not assume “accumulation” always works. In some fields, accumulation has exponential effects, but opportunity windows (like the AI wave) and the rising cost of trial with age are real. If potential energy is not released, it will eventually depreciate into a heavy sense of “unfulfilled talent.”

5. Action Design: Create Structural Change

If you do not actively create structural change, change will not happen by itself. In 2026, commit to gradual restructuring in three dimensions:

1. Time structure change

Goal: time you truly control.
Action: allocate X fixed hours each week to work that does not depend on a single employment relationship. This is not “spare time”; it is “seed time.”

2. Income structure change

Goal: not dependent on a single employment relationship.
Action: set a concrete milestone such as “side income reaches Y% of main income.” You can even set a required micro-commercial loop: even if you earn just one dollar, run the whole process end-to-end.

3. Identity structure change

Goal: how others introduce you to a third party.
Action: in certain circles, proactively introduce yourself with a new label. No longer “the technical lead of a system,” but “[your new identity].”

4. A concrete test

Action: choose one project, set a public deadline, and introduce external supervision (or prepaid cost). Make “preparation” itself a public, binding action.

Conclusion: This Is a Choice

This essay is itself a curious paradox: you are using “build a complete cognitive system” to criticize the pattern of “using systems to delay action.”

So let this summary be the last time you only “build cognition.”

This is not to create anxiety, but to make time feel heavy again. You are not lost. You are more like someone standing at a crossroads, unwilling to admit that you already see the direction.

How long will you stand here? That itself is a choice.

Zootopia 2: The Cost of Harmony?

Mon, 15 Dec 2025 02:47:05 GMT

From Fairy Tale to Allegory

When children laugh at Judy and Nick’s adventures in the theater, adult viewers should realize that Zootopia 2 is far more than a family-friendly animation. The sequel continues the tradition of wrapping complex sociological questions in a colorful animal world, but its critical edge goes deeper.

If the 2016 Zootopia exposed personal prejudice and hidden discrimination, the 2025 sequel turns its camera on a more fundamental question: how oppression is woven into social structures themselves. This is no longer a moral fable about “good people can be biased too,” but a systematic critique of colonial history, spatial politics, and inherited power.

The English title Zootopia is a carefully designed paradox: “Zoo” (cage) plus “Utopia” (ideal society). It hints that the city, though seemingly perfect, is built on exclusion and control. The sequel’s job is to tear away the utopian mask and show the violent foundations beneath the prosperity.

Revealing Three Layers of Structural Oppression

The most important theoretical contribution of Zootopia 2 is its presentation of oppression as a three-layer structure, moving from the surface down to the roots.

Personal-Level Bias

This was the focus of the first film and the starting point of the sequel. Judy carrying fox repellent spray and blurting out offensive remarks to Nick reveals the mechanics of implicit bias and microaggressions. Even people who see themselves as progressive can unconsciously reproduce discriminatory logic.

In the sequel, that personal bias extends to fear of reptiles. Gary, a gentle and even timid snake, stands in stark contrast to the city’s image of reptiles as “cold-blooded killers.” This contrast is no longer about biology; it is pure cultural construction. Reptiles are demonized simply because they are “other.”

Institutional Discrimination

The police system itself becomes an object of scrutiny. Judy, as the “first bunny officer,” experiences not just tokenism but a broader mechanism of exclusion. The mandatory “partner counseling” for Judy and Nick looks like therapy on the surface, but in practice is discipline imposed by the institution on nonconforming behavior.

A deeper critique is that the film hints at the police as guardians of power. Judy and Nick may be heroes, but they operate inside the system and rely on state violence to “save” oppressed groups. This extends the critique of the first film: a “white savior” and “blue lives matter” narrative that still places hope in establishment goodwill rather than empowering the oppressed to resist.

Structural and Systemic Oppression

This is the most subversive part of the sequel. It reveals that discrimination is not only in people’s minds or in written policies, but solidified in the city’s physical space itself.

The climate wall system, the segregation of Marsh Market, and the historical absence of reptiles all point to one core proposition: Zootopia’s prosperity rests on the systematic exclusion of certain groups. This is not a conspiracy by a single politician, but the original violence embedded in the city’s construction.

Metaphors

A. The Absence and Return of Reptiles

The complete absence of reptiles in the first film is revealed as a historical expulsion, directly echoing the logic of settler colonialism:

First, stigmatize the indigenous. A manipulated murder case (a snake killing a turtle) labels the whole group as “dangerous” and “savage.”
Second, physical expulsion. Reptiles are driven out of the city core and their community, Reptile Ravine, is destroyed or repurposed.
Third, historical erasure. Their contribution to the city (inventing climate wall technology) is erased, and the credit is given to the colonial Lynxley family.
Finally, a legitimizing narrative. The city’s “founding myth” casts the founders as heroes and justifies the expulsion as a matter of “security.”

B. The Climate Wall System

In the first film, the climate wall is framed as a technological miracle that allows species with different climate needs to coexist. The sequel flips this optimistic reading:

The climate wall is not a tool of inclusion, but a weapon of segregation.
It divides the city into climate zones, a physical embodiment of spatial separation.
The design of arid and cold zones deliberately excludes reptiles’ survival needs, a form of “environmental violence.”
Control of the wall’s maintenance and expansion sits with the Lynxley family, making it a tool to monopolize resources and control movement.

C. Marsh Market

Marsh Market is the most symbolic setting in the sequel. Built on wetlands and connected by boardwalks and pipes, it carries a bayou slum vibe. It is Zootopia’s “reverse side”:

Spatial marginalization: cut off from the city core, it becomes a gathering place for reptiles, aquatic mammals, and other “unwanted” groups.
Economic marginalization: smuggling and black markets flourish, not because of moral failure but because residents are excluded from the legal economy.
Cultural marginalization: the area has its own vitality and community culture, yet is seen by the mainstream as “dangerous” and “chaotic.”

The Lynxley family plans to expand Tundratown, which means swallowing Marsh Market. This is a blunt metaphor for gentrification:

Under the banner of “urban development,” low-income communities are displaced, land is converted into high-value real estate, and original residents are left with nowhere to go.
The “independence” of Marsh Market residents is essentially abandonment and exclusion, not true autonomy.

D. The Lynxley Family

The Lynxley family is the perfect embodiment of elite power:

Stealing innovation: climate wall technology was invented by reptiles, yet the Lynxley family claims it and patents it. This reveals the violent nature of intellectual property under capitalism: whoever controls law and power can legitimize others’ creations as their own.
Historical revision: by controlling education and public narratives, they cast themselves as “city builders” and erase reptiles from history.
Inherited power: wealth and influence pass down generations, forming a fixed ruling class. Even the “fallen” heir Bobert can easily mobilize smuggling networks and antique resources.

From “We Are the Same” to “Accepting Our Differences”

This thematic shift exposes the inner dilemma of liberal pluralism.

The First Film’s Optimism: The Assimilation Illusion

The slogan “Anyone can be anything” is essentially an assimilationist narrative. It assumes that once prejudice is removed, everyone can gain equal opportunity within the existing system.

Judy becoming a cop and Nick “going straight” both happen by adapting to and integrating into the mainstream system. This narrative ignores a fundamental question: what if the system itself is the source of oppression?

The Sequel’s Compromise: The Cost of Acknowledging Differences

The sequel retreats from “we are the same” to “accept our differences.” It seems more mature, but actually reveals a deep sense of powerlessness:

Biological differences are not easily changed: foxes do not understand hugs, elephants fear mice, and herbivores’ smiles look like provocation to predators. These details suggest some differences are deeply rooted and cannot be erased through “mutual understanding.”
Limits of inclusion: Judy and Nick’s painful faces when eating insects to respect reptiles, and seals refusing coins and touch yet still being ignored, show how difficult true cross-cultural understanding is.
The nature of compromise: “accepting differences” often means marginalized groups must adapt and sacrifice. In the film, reptiles work to fit into Zootopia rather than the city changing itself to welcome them.

This shift is, in effect, a critique of multiculturalism as ideology. It shows that “tolerance” and “respect” alone cannot solve structural inequality. Real justice requires a fundamental redistribution of power, not just the acknowledgment of difference.

The Irony of the Name Zootopia

The name itself is a precise critical tool.

The Double Meaning of “Zoo”

A zoo: a cage for humans to observe animals, where animals are classified, displayed, and controlled.
A city of animals: seemingly free, yet planned, monitored, and managed.

The False Promise of “Utopia”

When Thomas More coined “Utopia,” the etymology already carried a double meaning: “eu-topia” (good place) and “ou-topia” (no place). Zootopia is exactly such a “no-place” good place. It promises equality and inclusion, but that promise rests on systematic exclusion of certain groups.

The Collapse of the “Melting Pot” Myth

America has long styled itself as a “melting pot” where cultures blend. That narrative hides a harsh reality:

Integration equals assimilation: minorities must give up their cultural identity and adopt the mainstream.
Selective inclusion: only those who “fit the standard” are accepted, while others are expelled or marginalized.
Ongoing segregation: real-world racial segregation and economic stratification never disappear, they just take more hidden forms.

Zootopia’s surface prosperity is built on Marsh Market’s poverty, the absence of reptiles, and the Lynxley family’s monopoly. This utopia is, at its core, an exclusive space of privilege.

The ending scene, with different groups peacefully making burgers together, is warm. But the line “territorial animals always feel the urge to expand” tells the truth: power struggles are permanent, and harmony is temporary.

A Sociology Textbook in Fairy-Tale Clothing

Zootopia 2 uses animal allegory to weave settler colonialism, racist urban planning, and intellectual property violence into what looks like a light adventure.

The most unsettling part is not what it reveals, but that after revealing it, the film cannot provide answers. The final harmonious image is fragile because power structures remain unchanged, economic inequality persists, and spatial segregation continues. That line about territorial animals is the film’s most honest moment: it admits there may be no perfect solution.

Ultimately, Zootopia 2 matters because it asks the right questions even if it cannot offer radical answers. It invites viewers to think rather than feed them conclusions. It acknowledges complexity instead of selling easy hope.

In an era that increasingly rejects complexity, that is a radical stance in itself.

When children grow up and rewatch this film, they will realize that the scenes that once made them laugh were actually a prophecy of the world they will face. The answers that were not spoken are waiting for their generation to find.

Author’s note: personal opinion, for reference only.

What Has AI Changed for Me?

Tue, 02 Dec 2025 05:33:05 GMT

Introduction: A Gentle Singularity

Over the past two years, I have felt like I was standing on a strange boundary: on one side is the familiar old world, driven by experience, manpower, processes, and projects; on the other is a new world accelerating forward, with AI seeping into every crack like groundwater.

Every quarter I analyze AI news, trends, and products, and I binge AI-related podcasts during my commute. That is just the environment I live in.

Yesterday I saw a video from Tim that opened with a chilling experiment: “draw a picture of a person drowning in the deep sea, struggling to breathe.” Nine human artists spent an hour on it, while six AI models took only 15 seconds. The harsher truth was that the AI images were high quality, highly varied in style, and could be regenerated endlessly.

That contrast sharpened a question that has been bothering me for a long time: is AI a wave, or a lighthouse? When AI can replace an hour of my effort in 15 seconds, what is the meaning of my existence? Is it a tool that changes the outside world, or a program that rewrites me?

Looking back, the changes AI brought me can be split into three layers: a structural rewrite of how I work, an exponential amplification of my capabilities, and a quiet shift in my life trajectory.

AI Turned Me from a “Task Doer” into a “System Architect”

In the past, writing an architecture document was a war of willpower. Words, diagrams, structure, details, every part had to be built from scratch.

Now the process feels like collaborative engineering: I propose the core framework and AI expands the structure; I add business context and AI fills in the logic chain; I identify risks and AI offers alternative paths.

A document that used to take at least three days can now take just one night. The quality is often higher because I can spend more energy polishing concepts rather than grinding through sentences.

Tim’s video offers an even harsher footnote: his team once spent an entire year making a special-effects shot for “Galaxy Train.” In a traditional workflow, that would take five professional VFX artists for more than half a year (around 500 hours). But in 2025, the AI model Veo3 generated the same train-through-the-galaxy scene in two minutes, even with sound effects.

This is the “paradox of effort”: the more time and labor we pour into pure execution, the less meaningful the outcome may be, because AI evolves faster than we can refine our craft. The software skills and tricks we used to study late into the night can be devalued instantly.

In short: AI lifted me from “executor” to “structural engineer.” I am no longer just finishing tasks; I am designing systems.

AI Expanded My “Personal Ability” into a Small Team

If human ability is like a motherboard, AI is a plug-in expansion card you can attach at any time.

When I write code, it becomes a pair programmer. When I write system plans, it becomes an architect. When I create content, it becomes an editor, a scriptwriter, a media library, and an idea engine. When I need to talk to a financial advisor, it instantly switches into a financial analyst.

That ability to “switch professions on demand” first shocked me, then made me dependent, and now feels calm and normal, as if I have a 24/7 “shadow team” always on standby.

More importantly, it made me redefine a single word: boundary.

Before, I would think, “I am not good at this, I probably cannot do it.” Now it becomes, “Let me try. With AI as a safety net, it will not be too bad.”

So I started trying things I used to avoid: rebuilding my blog, creating content, understanding macroeconomics, preparing learning paths for my child. I began to connect across multiple domains and no longer feared crossing boundaries.

This is not inflation. It is a world quietly opened.

AI Helped Me Reshape the “Meaning of Work”

This might be the deepest change.

For a long time, work felt alienating: trivial, fragmented, passive, like a gear inside a machine.

But once I started using AI, I realized I was turning myself from a gear into a system designer. That shift itself is meaning.

When I design a two-year system evolution plan for my company, I am using AI to pull what I wanted to do but could not, back into my capability range.

When I design a learning path for my child, I am using AI to turn a father’s anxiety into deliberate action.

When I keep writing, I am using AI to turn accumulated experience into reusable knowledge assets, instead of scattered, temporary skills.

In other words, AI let me see that many things in work and life were not “I am not capable” but “I lack the right tools.” Once the tools arrived, my participation in the world was reactivated.

But that reactivation is not only about efficiency, it is about connection. Tim’s video told the story of the band “Deep Sea Breathing” and its singer, 57, which gave me a deeper understanding of “the meaning of work.”

57 had a serious illness, lost hearing, and faced a surgery with only a 60-70% success rate. He said something that broke my heart: “I am not afraid of life ending early. I am afraid that one day I will not be able to hear anything.” To help him fulfill his dream, the team rushed to finish the “Galaxy Train” MV before he went into surgery.

The surgery did not go perfectly. There was no miracle. He lost 30 kilograms, and his hearing did not improve much. But on the day the MV was released, 57 successfully proposed to his partner. In that moment, the efficiency of production and the precision of special effects (things AI can easily surpass) did not matter. What mattered was the bond between people that the project carried.

The meaning of work is not only about designing systems or boosting efficiency. It is about whether, through those systems, we can transmit love and courage. AI can generate perfect images, but it cannot understand the nervous sweat of a proposal, nor recreate the smell of dreams in a 60-yuan rehearsal room.

So What Do We Still Have?

If AI can do everything, faster, stronger, smarter than us, what do we have left?

Tim’s video gave a soul-piercing answer: “Maybe only experience. Experience shapes each of our unique souls.”

When you see your mother’s gray hair, your thoughts do not stop there. When you step onto your old campus and smell the classrooms, your memory is instantly awakened. These moments of feeling and these unique life experiences cannot be calculated or simulated by AI. AI cannot see through you, because it has not lived your life.

The final metaphor in the video about a “train” is deeply moving: in AI’s logic, trains are slow, heavy, inefficient, nowhere near fiber optics or rockets. It may not understand why humans feel attached to such an outdated vehicle. But trains have something special: “Even if boarding means separation, as long as the tracks remain and we still remember each other, one day we will meet again.”

Every time I finish overtime and drive home, I choose Gulou Avenue, even though the Second Ring Road is faster late at night. Besides speed, there are other things that matter.

This helped me see the other side of the AI revolution: AI can be the rocket that roars past, taking us higher to explore the boundaries of cognition. But we are still the person sitting by the train window (or driving), holding the hand of someone beside us and feeling our heartbeat.

Efficiency belongs to AI. Experience belongs to humans.

The Value of AI Is Never Replacement, but Upgrade

It brings efficiency, but also direction. It changes work, but also changes mindset. It accelerates professional growth, but also reshapes self-understanding.

I do not know what the future holds, but at least right now I can feel one thing clearly: AI is not just a tech trend, it is a personal revolution. It forces us to hand over inefficient labor in exchange for more precious experiences and connections, and I happen to be in the center of that revolution.

Author’s note: personal opinion, for reference only.

What Is a Tariff? Why Is Trump Raising Tariffs on China Again?

Sun, 27 Apr 2025 09:11:11 GMT

I recently talked with a friend who does import-export business in the US. He has already paused imports from China and switched to transshipment or sourcing from other countries. His reason was simple: the US raised tariffs again, margins got crushed, and the business no longer works.

Influenced by that, I rushed to order a new Mac. I was afraid that if I waited and the tariffs landed, prices would rise and consumers like us would pay the price.

That sparked my curiosity: what are tariffs really? Why does the US like to talk about them? How much do they affect ordinary people?

Over the past few weeks I did some research and put together this beginner-friendly guide to help explain the logic behind tariffs.

1. What Is a Tariff? From “Toll” to Geopolitical Weapon

1. What is a tariff?

A tariff is essentially a tax imposed by the government on goods that cross borders. You can think of it as a modern “toll” or “protection fee.”

For example, if you buy a pair of sneakers from abroad, when the goods enter Chinese customs, the state collects a tax based on the declared price. That tax is a tariff.

2. What are its three main functions?

Function	Explanation
Fiscal revenue	A key source of revenue for early states; still important for developing countries
Trade protection	Raises foreign goods prices to protect domestic industries
Political bargaining tool	Used to threaten or sanction other countries, e.g., the Trump-era “trade war”

In modern economies, tariffs are no longer just “taxes” but comprehensive national policy tools with strong strategic attributes.

2. The US and Tariffs: A Tangled History

1. Tariffs are an “old friend” of the US

In the early years of the US, the federal government had no personal income tax; almost all fiscal revenue came from tariffs.
One trigger of the Civil War was the South’s opposition to high tariffs set by the North (protecting Northern industry and hurting Southern agriculture).
So US reliance on tariffs did not begin with Trump; it is “in the family.”

2. Trump: Governing with tariffs

Starting in 2018, the Trump administration launched four rounds of tariff actions against China.
Covered more than $360 billion of Chinese exports.
Average US tariffs on Chinese goods rose from 3.1% to 21% (source: Peterson Institute for International Economics, 2020).

3. Why can he “raise them at will”?

US law gives the president many unilateral tools to raise tariffs:

Legal provision	Allows tariffs on the grounds of…
Section 301	Retaliatory tariffs against trading partners
Section 232	”National security” tariffs (e.g., steel, aluminum)
IEEPA	In emergencies, tariffs, asset freezes, embargoes (used against fentanyl)

3. A New Tariff Storm: Is It Just “Trump Showmanship”?

During the 2024 campaign, Trump proposed more aggressive tariff plans:

Up to 125% tariffs on Chinese goods
25% tariffs on neighbors like Canada and Mexico, citing “drugs and illegal immigration”
A global 10% baseline tariff, because “other countries take advantage”

These statements are partly campaign theatrics, but they also reflect that the US is treating tariffs as a normal national policy tool.

4. Why Does the US Keep Raising Tariffs? Surface vs Deep Logic

Surface-level reasons (official narrative)

Reason	Example
Trade deficit is too large	China exports far more to the US than the US exports to China
Protect manufacturing jobs	Promote “manufacturing reshoring”
National security	Sensitive areas like chips, rare earths, fentanyl

Deeper layer: the US “triple dilemma”

1. Triffin dilemma: side effects of dollar hegemony

As the issuer of the “world currency,” the US must continuously export dollars (i.e., buy from other countries) to meet global dollar demand.
This leads to long-term trade deficits and hollowing-out of manufacturing.

To maintain dollar hegemony, manufacturing must be sacrificed. This is the Triffin dilemma.

2. Rust Belt votes and populist demand

Traditional manufacturing regions in the US Midwest have long faced high unemployment and economic stagnation.
Trump’s base is working-class voters in these “Rust Belt” areas.
“Tariffs on China” can trigger their sense of deprivation + nationalist emotions. It is a powerful political tool.

3. High fiscal deficits, need “hidden taxation”

US national debt has exceeded $34 trillion (as of end-2024, source: US Treasury).
Tariffs are not just trade policy but also a way to tax domestic consumers, increasing fiscal revenue without directly raising income taxes.

5. Do Tariffs Work? Highly Contested

Pro arguments	Con arguments
Short-term protection for manufacturing	Distorts price mechanisms and reduces resource allocation efficiency
Increase government revenue	Burdens companies and consumers
A bargaining tool	Disrupts global supply chain cooperation

According to a 2021 report from the US Congressional Budget Office (CBO), 93% of the tariff cost in the last trade war was borne by US firms and consumers, not Chinese exporters.

6. Who Pays for Tariffs? Consumers and Companies

Although factors like RMB depreciation can offset some impact in the short term, in the long run:

US consumers bear most of the cost through higher prices;
Companies in both China and the US must sacrifice on margins, capacity, and supply chains;
Global trade becomes more uncertain, accelerating supply-chain “de-Chinaization.”

Tariffs are a weapon that “seems to hit others but actually hits yourself and the global market.”

7. Summary

The US keeps wielding tariffs. On the surface, it attacks opponents, but in essence it tries to use power to respond to its structural problems:

Side effects of dollar hegemony
Industrial hollowing and populism
High deficits and debt-driven fiscal imbalance

In this process, China is simply the most convenient “imagined enemy.”

No matter who becomes president in the future, tariffs will only be used more often and will not disappear. This is a reflection of global “de-globalization” and geopolitical fragmentation.

Appendix | A Simple Guide to Tariffs, the Dollar System, and US Debt

Tariff terms

Term	Simple explanation
Tariff	A tax on imported goods
Baseline tariff	The default rate
Reciprocal tariff	A matching response to another country’s tariff policy

What is SWIFT?

A “messaging system” for secure transfers between banks.
It does not move money itself, only sends instructions.
Analogy: the “logistics network” of banking.

The Bretton Woods System

A global monetary system established in 1944, with the dollar pegged to gold and other currencies pegged to the dollar.
It laid the foundation for dollar hegemony.

Triffin Dilemma

The world wants dollars, so the US must keep exporting dollars (buy, buy, buy).
But too many dollars weakens trust.
The contradiction between “globalizing the dollar” and “US self-interest.”

The Global Role of US Treasuries

Absorb excess dollars globally, serving as the credit anchor of the dollar.
Provide low-cost financing for the US government.
Form a loop with the dollar: “dollars -> buy Treasuries -> the US spends again”

As you can see, tariffs are far from a simple “economic issue.” They are a node where politics, fiscal policy, currency, and international order intertwine.

How Small Teams Implement Agile Development

Mon, 14 Apr 2025 06:29:01 GMT

You can’t manage what you don’t measure. – Peter Drucker
If you can’t measure it, you can’t manage it.

This is a famous quote from Peter Drucker, the father of modern management. For project management and agile development, the prerequisite is to connect the data and make it visible and measurable—only then can you see it and manage it.

This article uses a SaaS product as an example to explain how a “small team” can implement agile development in practice.

Why Implement It

Progress is opaque; we don’t know where the work actually is.
Delayed releases become the norm, and we don’t know when things will ship.
After releases, we feel uneasy and don’t know when issues or failures will appear.
Communication cost is too high, and RD/FE/QA/PM are frequently misaligned.
Requirements are inserted arbitrarily and frequently, without cost awareness.
It’s unclear whether the R&D workload is normal, overloaded, or underutilized.

In internet startups, demand and limited resources are always in conflict. Finding balance, focusing limited resources on strategic needs, and keeping team rhythm and morale are the pain points that agile management aims to solve. It also provides answers to the problems above.

Tools We Use

The tools you use shape the way you work.

Atlassian Confluence
Atlassian Jira

How to Do It Well

“Requirement Review ➡️ Design Review ➡️ Development ➡️ Testing ➡️ Acceptance ➡️ Release ➡️ Post-Assessment”

To make product and R&D processes visible, controllable, and synchronized, we adopted a four-board agile management model. The boards and their functions are as follows:

Public Requirement Board (Kanban Board)

Use this board to build a public requirement pool and collect requests from cross-functional stakeholders. Market feedback is routed here promptly. Board type: Kanban.

Requirement Board (Kanban Board)

Build a lifecycle for requirements and visualize flow across stages: “Backlog – Review – Scheduling – Design – Development – Release.”

Task Efficiency Board (Scrum Board)

Predefine release windows so everyone can see schedule risk early. Product, engineering, and requesters can communicate quickly to sync changes or progress.

Bug Board (Kanban Board)

Track bug counts by type within the iteration clearly.

Public Requirement Management (Public Requirement Board)

Our company is in the education SaaS space. Public requirements mainly come from:

Key customers (schools)
Daily user feedback (teachers, students, parents)
Sales channels

We identify and filter fake, minor, or non-strategic requests here. Managing the many requests from business stakeholders is still hard. Common sources include:

Sales
Marketing
Internal (product roadmap / other)
User experience feedback

Customer success, sales, and other stakeholders can submit public requirements on this board. Product managers filter and research them, then convert them into product requirements and move them into the product backlog. The Public Requirement Board card fields typically include: requirement description, source, reporter, and owner.

If a requirement has very low value or will definitely not be done, move it directly to Done.
If it has some value or needs analysis, move it to Research & Discussion. After confirmation, move it to Done.

Product R&D Requirement Management (Requirement Board)

Requirement categories

Internally, we split requirements into two types:

Product requirements: iteration, urgent, and routine requirements proposed by PMs
Technical requirements: refactoring and improvement work for stability, scalability, and maintainability

Requirement priority

An old saying goes: “An army must have a justification.” Requirements should, too. To help engineers understand importance and urgency, priority levels are essential.

Product requirement levels
- P0: critical tasks that must be completed as fast as possible; can reassign people, pause other projects, and require overtime
- P1: very important tasks with a deadline; cannot be delayed; if a P0 appears, overtime is needed to protect the P1 deadline
- P2: important and impactful tasks with a deadline; can be postponed when P0/P1 appear (most tasks)
- P3: nice-to-have tasks, lowest priority
Technical requirement levels
- T0: major performance or security issues, requires urgent fixes
- T1: scalability or performance risks, usually handled as separate tasks
- T2: design or general performance issues, often improved during iterations

Product requirement management (Requirement Board)

PMs and engineers communicate through PRDs. Engineering ultimately develops and tests based on PRDs, so the first step is creating a PRD. PRDs are managed flexibly. Some PMs use Lanhu, others use MockingBot. To unify archiving, we store PRDs in Confluence (the detailed link can point to any tool). When creating in Confluence, use a template, as shown below:

Main contents include:

Background description
Business goals
Requirement links and feature list (stories)

The Requirement Board has four swimlanes: P0, P1, below P1, and Technical Requirements. To make searching easier, we added common keywords to quick search. Each card includes: priority level, due date, and owner.

Technical requirement management (Requirement Board)

Examples include data structure changes or architecture improvements (e.g., switching to Apollo for config). These are considered technical requirements. Their display and board behavior is basically the same as product requirements and also appears on the Requirement Board.

Technical Task Management (Task Efficiency Board)

This stage transitions from requirements to development. The tasks here include:

Development tasks: RD, FE
Developer self-testing: RD, FE
Test case writing: QA
Test case execution: QA

These tasks mainly come from two sources:

Product requirements
Technical requirements

We use the Task Efficiency Board to manage them. Before starting, drag the product or technical requirements to be iterated from the Backlog, as shown:

Then create sub-tasks under each product/technical requirement in the Requirement Board, as shown:

After creation, you can view parent/child task details and see effort tracking:

Start the sprint and set the timeline:

RD, QA, and FE fill in their daily work logs before the end of each day so we can track effort:

Bug Management (Bug Board)

Bugs created during a sprint appear on the Bug Board. The workflow is: Open ➡️ In Progress ➡️ Resolved ➡️ Closed.

We classify bug types as:

Functional defects
UI improvements
Feature improvements
Performance issues
Security-related issues

After the iteration ends, QA analyzes bug statistics and produces a test report. We created a unified Confluence template for test reports. The main contents look like this:

Summary

Lifecycle management for requirements and efficiency is customized to our product and team stage. It should continue evolving as people and phases change. Below are the core processes and methods we use in Jira and Confluence:

4 boards
- Public Requirement Board: handles “key customer demands” and “user experience issues” from market/sales-facing teams
- Requirement Board: manages product and technical requirements
- Task Efficiency Board: manages development phase tasks and effort tracking across RD/FE/QA
- Bug Board: manages five bug categories within iterations: feature improvements, functional defects, UI improvements, performance issues, and security-related issues
2 templates
- Product requirement template: manages product requirements
- Test report template: summarizes bugs and testing after each iteration

At this stage, we’ve only implemented a small part of agile management. Many more practices will be added and refined over time.

Agile management exists to deliver products quickly and reliably. Don’t let the pursuit of agile tools distract you from the real goals.

Git Commit Guidelines

Mon, 14 Apr 2025 06:14:55 GMT

Why We Need It

Reduce review cost by making changes and impact obvious.
Standardize the team’s commit habits and improve engineering discipline.
Use a consistent format so tools can generate changelogs automatically.

GitHub Angular Demo

The Angular project on GitHub follows a strict Git message convention for daily commits and release management. Below is a snapshot of its commit history and the changelog generated from those commits.

Which Standard to Follow

The most commonly used convention is the AngularJS standard.

angular/angular.js

# Includes three parts: Header, Body, and Footer

<type>(<scope>): <subject>
// blank line
<body>
// blank line
<footer>

Header includes three fields: type (required), scope (optional), and subject (required).

No line should exceed 100 characters.

type

Type describes the commit category. Common types include:

feat: A new feature
fix: A bug fix
docs: Documentation only changes
style: Changes that do not affect the meaning of the code (white-space, formatting, missing semi-colons, etc)
refactor: A code change that neither fixes a bug nor adds a feature
perf: A code change that improves performance
test: Adding missing or correcting existing tests
chore: Changes to the build process or auxiliary tools and libraries such as documentation generation
revert: Reverts a previous commit
build: Changes that affect the build system or external dependencies (example scopes: gulp, broccoli, npm)
ci: Changes to our CI configuration files and scripts (example scopes: Travis, Circle, BrowserStack, SauceLabs)

If the type is feat or fix, the commit will appear in the changelog. For other types, decide whether they should be included.

scope

Scope describes the area affected, such as data layer, controller, or view layer. It varies by project.

subject

Subject is a short description of the commit’s purpose.

Body

Body is a detailed description of the commit and can be multi-line.

Footer is used only in two cases.

Breaking changes

If the current code is incompatible with the previous version, start the Footer with BREAKING CHANGE, followed by a description, rationale, and migration steps.

Closing issues

If the commit targets an issue, you can close it in the Footer.

Example: Closes #123, #245, #992

Tooling Constraints

Our goal is to generate and enforce conventions through tools.

Commitizen

commitizen/cz-cli replaces git commit.

Use the git cz command to generate commit messages that follow the convention.

# Install (make sure npm is installed first)
# Global install of commitizen
npm install commitizen -g

You also need to specify an adapter for Commitizen, such as cz-conventional-changelog (a preset that follows the Angular convention), so Commitizen can generate compliant messages.

# Go to your repo root

cd your_repo_root_path
# Initialize package.json
npm init --yes
# Set the adapter for commitizen
commitizen init cz-conventional-changelog --save-dev --save-exact

Now you can use git cz for commits.

However, if you still use git commit -m "", Git will accept it, which isn’t what we want. We need format enforcement, so we add commitlint + Husky.

Commitlint + Husky

commitlint checks commit message format.

If a commit does not match the target convention, it should be rejected.

The best way to validate commit messages is via git hooks, so we use Husky.

# In your repo root
npm install --save-dev @commitlint/{config-conventional,cli}
npm install husky --save-dev

# Add to the end of package.json
"husky": {
    "hooks": {
      "commit-msg": "commitlint -E HUSKY_GIT_PARAMS"
    } 
}

# Create commitlint.config.js in the repo root
module.exports = {
  extends: ['@commitlint/config-conventional'],
  rules: {
  'type-enum': [2, 'always', [
     "feat", "fix", "docs", "style", "refactor", "perf", "test", "build", "ci", "chore", "revert"
   ]],
  'scope-empty': [2, 'never'],
  'subject-full-stop': [0, 'never'],
  'subject-case': [0, 'never']
  }
};

Now try git commit -m "test" and you should see it blocked:

Standard Version

With the tooling above, your commit messages should follow the Angular convention. That also makes it easy to use a tool like standard-version to generate CHANGELOGs and even semantic versions.

# In your repo root
npm install --save-dev standard-version
# Add a script to package.json
{
  "scripts": {
    "release": "standard-version"
  }
}
# Generate changelog
# First time
npm run release -- --first-release
# Later
npm run release

A CHANGELOG.md file will be generated in the repo root:

How to Use It in a Project

After completing the steps above, you’ll end up with a package.json. Commit only package.json and commitlint.config.js to your repo.

Add node_modules and package-lock.json to .gitignore.

When someone clones the project, they can run npm install in the repo root to set everything up.

PS: If npm mirrors are unstable overseas, you can switch to the Taobao registry.

npm config set registry https://registry.npm.taobao.org

Git Branch Guidelines

Mon, 14 Apr 2025 06:11:02 GMT

Problems We Face in Day-to-Day Development

When we need to hotfix a production bug, how should we branch? Do we change develop or master directly?
We often have multiple parallel features with different release dates, owned by different sub-teams. Managing these parallel tracks and shipping the correct code is hard.
Everyone names branches differently, and interpretations vary. Right before release we discover the wrong branches were merged.

Goals of Branch Management

Commit code to the branch where it belongs.
Be able to switch to the stable production code at any time.
Develop multiple versions in parallel.
Make the purpose of each branch clear and ensure operations match that purpose.
Use Git to manage iterations, hotfixes, and feature development.

A Reference From a Predecessor

A successful Git branching model
By Vincent Driessen on Tuesday, January 05, 2010

A successful Git branching model

This is not official Git or GitHub documentation. It’s a personal summary based on that author’s team at the time, so it doesn’t map directly to our current workflow. After studying Git Flow, we defined our own team conventions.

Our Conventions

feature/sprintXX

Purpose: new feature iteration development
Environment: DEV
Create: from develop
Delete: after release
Restriction: none

test/sprintXX

Purpose: testing for new feature iterations
Environment: FAT
Create: from feature branch (before each test, merge develop into feature, then merge feature into test)
Delete: after release
Restriction: No Push / developer merge

hotfix/yyyyMMdd

Purpose: production issue fix
Environment: FAT
Create: from master (after release, merge back to master and develop)
Delete: after release
Restriction: none

develop

Purpose: contains the latest feature code, base for feature/sprintXX
Restriction: No Push / developer merge

master

Purpose: stable code
Environment: PROD
Restriction: No Push / merge by limited maintainers

Branch Operations

Use rebase on `feature/sprintXX` branches

To keep a clean commit graph. git pull defaults to merge, so you can use rebase:

git pull --rebase

# also set a global config
git config --global pull.rebase true
git config --global branch.autoSetupRebase always

Use `no-ff` for merges

Fast-forward merges produce a straight line and hide the branch history. Using --no-ff makes feature branch merges explicit:

# Merge sprint01 into develop
git merge --no-ff feature/sprint01

Code Review Practice

Mon, 14 Apr 2025 05:52:08 GMT

What Is Code Review

Google Code Review’s description

The biggest thing that makes Google’s code so good is simple: code review. At Google, no code, for any product, for any project, gets checked in until it gets a positive review.

The reason Google’s code is so good is simple: code review. At Google, no code for any product or project gets merged without review.

It is a culture, not a policy.

Why It’s Needed

Pain Points at Work

Everyone’s code style is inconsistent
Interfaces, classes, and names are inconsistent
Deep logic problems are hard to catch via tests
Complaints like “Who designed this? So many traps (maybe it was me months ago)”

What It Brings

Better code quality
Team knowledge sharing
Unified team standards
A shared culture of good code

What to Review

Google Review Principles

What to look for in a code review

Design: Is the code well-designed and appropriate for your system?
Functionality: Does the code behave as the author likely intended? Is the way the code behaves good for its users?
Complexity: Could the code be made simpler? Would another developer be able to easily understand and use this code when they come across it in the future?
Tests: Does the code have correct and well-designed automated tests?
Naming: Did the developer choose clear names for variables, classes, methods, etc.?
Comments: Are the comments clear and useful?
Style: Does the code follow our style guides?
Documentation: Did the developer also update relevant documentation?

Our Review Principles

Architecture and Design

S.O.L.I.D
- Single Responsibility Principle (SRP)
- Open/Closed Principle (OCP)
- Liskov Substitution Principle (LSP)
- Interface Segregation Principle (ISP)
- Dependency Inversion Principle (DIP)
Consistent behavior
- Is caching consistent? Is error handling consistent? Are error messages consistent?
- Do the same logic/behavior go through the same code path? Another symptom of low-quality code is that the same behavior/logic appears in different places or is triggered in different ways, without a unified code path, or with copy-pasted implementations everywhere, making maintenance very hard.
DRY principle (Don’t repeat yourself)
Program to interfaces, not implementations
Robustness
- Are corner cases fully considered? Is the logic robust? Any potential bugs?
- Any memory leaks? Any circular dependencies (language-specific, e.g., Objective‑C)? Any wild pointers?
- Thread safety? Data access consistency?
Error handling
- Good error handling? e.g., network errors, IO errors
Efficiency / performance
- For frequent messages or large data, are time‑consuming operations handled properly?
- What’s the time complexity of critical algorithms? Any potential performance bottlenecks?

Style

Engineering conventions (project structure, layering, naming, etc.)
Apollo configuration conventions
Naming conventions (interfaces, classes, methods, variables, etc.)
Code style (braces, spaces, line breaks, indentation)
Comment conventions (required comments)
Logging conventions (log what’s necessary)
DB SQL conventions
URL conventions

How to Do Code Review

Team Prerequisites

Git Branch Model
Git commit messages (should clearly describe changes)

When to Trigger

Tools

GitLab
SonarQube

References

As an R&D Leader, How to Summarize

Mon, 14 Apr 2025 03:58:33 GMT

Why Summarize

“When actions fail, seek the cause in yourself.” — Mencius

Learning to look for reasons in ourselves can help us become better people.

Whether for individuals or teams, doing a targeted year-end summary (preferably data-driven) is very helpful for finding problems and improving.

Avoid summarizing just for the sake of summarizing. We should use objective facts (data is a good tool) to find the causes behind problems, then try to change and improve.

Preparation

A thinking framework is more important than the thinking content.

Avoid a流水账

As an R&D leader, when summarizing I first avoid a流水账 (chronological diary). Don’t start by thinking “what we did this year,” because that easily gets lost in details.

Think: What Kind of Team Do We Need?

As a child, we were often asked, “What kind of person do you want to become?” Managing a team follows the same logic: first think about “what kind of team we need.” At the start of 2021, our guiding idea was “become a professional team,” and over the following year we adhered to that principle.

Guiding Principles for Team Building

Abstract the Team’s Work Categories

My Approach

What We Did

Recommend describing key items according to the work categories.

What We Achieved, What We Didn’t

Compare what we did with objective data feedback and the goals set at the beginning of the year. Analyze the gaps, find root causes, and provide improvement ideas for the next year.

The intent is to raise team cognition: giving a fish is not as good as teaching how to fish.

My 2021 Summary (Partial)

What We Did

Iteration

Arch
- Backend
  - Content too sensitive
- Frontend
  - Content too sensitive
Assets
- Tools
- Servers and environments
- Code planning & design
  - Content too sensitive
Team
- Architecture
  - Content too sensitive
- Incentives
  - Content too sensitive
- Activities
  - Content too sensitive

What We Achieved, What We Didn’t

Content too sensitive

As an R&D Leader, How to Plan

Mon, 14 Apr 2025 03:49:59 GMT

Real-World Planning

In the Q2 that just ended, we reviewed iteration results and OKRs, and then immediately had to break down Q3 OKRs. During that process, I talked with a few teammates and found that most people see planning as listing problems, setting some targets for each, and then executing. This makes target accuracy, logical consistency, and even whether something is truly a goal highly uncertain.

Methods of Planning

The way we plan reflects how we think when facing problems. Here we need to understand two basic logical analysis methods: deduction and induction.

Deduction

A method of reasoning that starts from a theory reflecting objective laws and infers unknown parts of a thing from its known parts.

Example

Aristotle’s classic syllogism is a form of deduction.

Major premise: All humans die. Minor premise: Socrates is human. Conclusion: Socrates will die.

If the premises are correct, the conclusion is correct.

Induction

A method of reasoning that, based on a series of experiences or knowledge materials, finds the basic or common laws they follow and assumes other similar things will follow those laws as well.

Example

Premise: My aunt made money with P2P, my younger aunt made money with P2P, coworker A made money with P2P. Conclusion: I will also make money with P2P.

This is a typical inductive inference: because we observe many people making money with P2P, we conclude that we can too, and we make a prediction about the future.

Which Method We Should Use

From the examples, it is clear that induction has risks: it treats the past performance of a thing as the thing’s inherent attribute, so it assumes the future trend will be the same.

With deduction, as long as the premises are correct, the conclusion is correct. Therefore, when we use deduction, we must ensure our major and minor premises are truly correct.

The method I use here is deduction.

The Value of Planning

Identify “Key Tasks”

The primary value of planning is to find the most important parts for the team or organization within complex reality, and then concentrate resources on execution.

Think from the Whole

Those who do not plan for the long term cannot plan for the short term; those who do not plan for the whole cannot plan for a part. — Chen Danran (Qing Dynasty)

Another value of planning is to think from the whole and from first principles, ensuring our plan is logical, reflects objective laws, and is based on the right premises.

As the definition of deduction says:

A method of reasoning that starts from a theory reflecting objective laws and infers unknown parts of a thing from its known parts.

Make Goals Everyone’s Goals

After planning, we should clearly explain the analytical reasoning and “why” behind the plan to the team. Only when everyone truly understands the plan can they execute the goals well.

A very important job of an R&D leader is to “identify key tasks,” find team and technical architecture problems, make systematic plans, and guide the team to solve them.

A 6-Month Planning Demo for a SaaS Product

R&D Management Practice with Structural Thinking

Mon, 14 Apr 2025 03:45:38 GMT

Why Think About “Structuring”

People

A few years ago, I was lucky to read The 7 Habits of Highly Effective People (search it on Douban if you are interested). It presents a crucial logic for how we live: “Begin with the end in mind.”

It tells us to be principle-centered rather than driven by events or appearances. We need stable principles that guide our lives.

Team

Team management is the same. As a leader, you need to identify the key tasks and keep the team moving in the right direction. So you also need to “begin with the end in mind”—find stable principles for an R&D team. Then output in a structured way, improving product capability and team capability around those principles and progressing continuously.

The Compass

The team’s “compass” guides work decisions.

We should follow the principle of “begin with the end in mind” to provide a stable, structured thinking framework and work framework for various choices and decisions.

That is the core meaning of the compass.

Composition

R&D Effectiveness Metrics

Goal: the ability to continuously deliver value quickly is the core objective of efficiency improvement.

Metric Item	Sub-metric	Description
Delivery throughput	Number of versions delivered per unit time	The number of versions delivered in 1 month or 1 quarter (trend-focused)
Demand response time	Delivery cycle	Average time from demand kickoff to release
-	Development cycle	Average time from coding start to production release
Delivery quality	Bugs per story point / function point	Calculated differently by iteration type
-	Online issues fixed per unit time	The number of online issues fixed in 1 month or 1 quarter
Continuous release capability	Release frequency	Number of releases in 1 month or 1 quarter
-	Release time	Average time from code merge completion to production release

System Availability Metrics

Goal: fully reflect system health, build SLI/SLO, and provide extremely high stability and availability to users.

Based on Google’s SRE theory, observe system health through SLI, SLO, and SLA.

Work Logic

Our team is responsible for a SaaS product, so we abstracted a structured model of our work logic.

R&D Capability Levels

Current levels: Poor, Fair, Average, Good

Takeaways

This abstract structured work logic has been running for a while and has produced some results, such as:

In planning, we now have a stable structure to continuously improve specific capabilities
In goals, the team is clear about the overall plan for the next two quarters
In decisions, we identify key tasks based on “structure” and “principles”

My 2024

Sat, 04 Jan 2025 09:18:33 GMT

Today is the first Saturday of 2025, and I am working overtime. I spent the day writing my team’s 2024 summary, and at night I finally had time to review what I did in 2024. If I pause and think, what would I say to my 2025 self?

At this moment, a famous line from Romain Rolland came to mind:

There is only one true heroism in the world: to see the world as it is and still love it.

For me in midlife, this line means something different: it sounds grand, yet feels ordinary. What is “life” anyway? The life each person has lived or can live is completely different. Standing at the start of 2025, I would tell my future self: build the ability to choose what kind of life you want. In a difficult environment, expand your capability boundaries and give the future more possibilities.

Travel

This year I visited a few places at home and abroad and experienced different cultures. I also met an old friend in Hangzhou and talked for two days straight.

February trip: Fuzhou -> Quanzhou -> Xiamen

Fuzhou: got our New Year’s Eve dinner before McDonald’s closed so we did not end up eating instant noodles on New Year’s Eve.

Quanzhou: half the city is fireworks, half the city is gods. There is a temple every few hundred meters. You can feel the ease and calm of living there, as if no problem is too big for so many deities.

Xiamen: a city I have visited many times. A classic seaside town. This time I skipped the famous attractions and just wandered.

March: visiting old friends in another city (Hangzhou)

I flew to Hangzhou to meet friends I had not seen in a long time. We talked about the past and the present. One takeaway: when friends live in different cities, meeting once often requires annual planning.

May trip: Ningbo

Dongqian Lake: beautiful scenery. I have been to Lugu Lake before; each has its charm.

Ningbo Museum: went mainly because it was a filming location for The Three-Body Problem.

July trip: Georgia

A very long flight: Beijing -> Urumqi -> Tbilisi (capital of Georgia), more than ten hours.

Simple folkways + European style + Asian prices: Georgia has kind people. Tourism is not very developed yet, so things are not expensive and there is no tourist trap vibe. My guide was local and spoke English, Chinese, Georgian, and Russian, so communication was easy.

Recommended places: Tbilisi city walk, Kazbegi, Batumi (Black Sea). The churches are so many I cannot remember them all.

September trip: Anaya Jinshanling

Went with a friend’s family. Not far (about 150 km from Beijing). Fully pet-friendly. The community feels very Anaya. Great for pet lovers and small group gatherings.

November: musical, Man of La Mancha

Took my kid to a live musical for the first time. I had not read Don Quixote systematically before, but I knew the theme. The show moved me and helped me rethink idealism and how to pursue it.

Work

Not much change. My team and I kept moving bricks and delivering. But in 2025, I need to keep independent thinking and focus on a few important goals.

Inputs and Outputs

Inputs

I adjusted some of my information intake tools this year. Here is what I use now.

Follow

Crushes every RSS tool I used before. Now my only RSS + newsletter tool.

Cubox

My long-term read-later tool for organizing scattered information.

Xiaoyuzhou

Listened for 138 hours this year. Two podcasts I cannot live without:

Between Reality and Fiction

The best podcast I have ever heard, no exaggeration. High quality and high cognition. Since it launched in 2023-10, I have listened to every episode.

Xiaoyuzhou - Between Reality and Fiction

69 hours total this year

Shengdong Morning Coffee

Every workday morning, 15 minutes on my commute to catch up on global tech and business news.

Xiaoyuzhou - Shengdong Morning Coffee

47 hours total this year

WeRead + Print Books

Zhang Juzheng

Historical novel. Zhang Juzheng was a formidable figure who pushed a series of reforms during the Wanli era, such as tax reform and cutting redundant officials, but it was extremely difficult. The book also shows how he navigated complex political struggles. You finish it thinking: how hard it is to be a reformer in history.

The Little Prince

Each journey is like a mirror that makes you reflect on friendship, love, and growth. On the surface it is a fairy tale, but it is really about life philosophy.

Rework: Escape the Madness of Busywork

Written by the founders of 37signals, it proposes a counter-mainstream approach to work and entrepreneurship, emphasizing less busyness, focus on core goals, and more effective execution. Think independently.

I and the Temple of Earth

Essays. The place became a stage for Shi Tiesheng to think about life. He writes about death, maternal love, and the meaning of life. The prose is plain but moving. The descriptions of the Temple of Earth are great material for kids’ essays.

Stories of the Sahara

Sanmao and her husband Jose lived in the Sahara Desert. The conditions were harsh, but their life was romantic. The book records interesting people and events and her love for life. It reads easily but has depth. I even searched Xiaohongshu to see photos of her former home, which made me feel more immersed.

A Brief History of Chips

In the year AIGC took off, I stumbled on this book. It brought back memories of my college courses and projects. From the first transistor to how chips shape global tech today, the author explains it clearly. You will see the tech story behind the scenes.

Inside the System

If you are interested in economics but find it too complex, this one is perfect. It explains how the Chinese economy operates in simple language and clarifies the difference between “inside” and “outside” the system. You will better understand our economic logic.

Beyond Sapiens

Harari’s new book. It explores how future AI and biotech may change humanity. It also reflects on the challenges of progress: are we happier, or have we lost something important?

DeDao

I took two history courses: Shi Zhan’s Outline of Chinese History, and Li Jun’s Outline of Western History. I love “big history” and these two courses are excellent. Recommended for anyone who likes wide historical narratives and connected timelines.

Duolingo

I kept learning English for 365 days again. But to keep more options open, I plan to increase my English effort in 2025. The goal is to reach a level where I can work and live in English-speaking countries (I am still far away, but a goal is a goal).

Keep going! Give myself and my family more possibilities. It is better to push myself than to push my kid; pushing kids is not very cost-effective right now.

Outputs

Notion

My main daily output tool for organizing ideas, managing work, and personal learning.

Blog

At the end of this year, I grabbed a cheap VPS during Black Friday (https://cloudcone.com/). After that impulse buy, the VPS could not sit idle, so I migrated my blog from a static Vercel site to a WordPress dynamic blog on the VPS.

I wrote only six posts this year, about one every two months. Next year, try harder.

Free Diagramming Tools for Engineers

Thu, 19 Dec 2024 09:21:51 GMT

Year-end is planning season again, and that always means diagrams. From architecture design to flowcharts (and even a few “magic” slides in meetings), you need a handful of diagrams to hold the room. But who says drawing has to cost money? If it can be free, it should be free. Here are a few free tools I use often.

Excalidraw - Hand-drawn creativity

http://excalidraw.com

If you find formal charts too stiff, try Excalidraw. Its hand-drawn style instantly gives you a “I am an artist” vibe, like Picasso drawing your system architecture.

Highlights

Collaboration support
Built-in asset library (add as needed)
End-to-end encryption; no user data stored

Mermaid - Syntax-driven, developer-friendly

https://mermaid.js.org

Tired of drag-and-drop? Mermaid lets you draw with code. It is the ultimate romance for engineers. When you generate a beautiful flowchart from text, it feels even better than fixing a bug.

Highlights

Generate diagrams from code, very hacker-style (AIGC makes it even faster)
Rich diagram types: flowcharts, Gantt, sequence diagrams, and more
Easy to integrate with Markdown editors

Drawio - All-rounder

https://www.drawio.com

If Excalidraw is the cute neighbor and Mermaid is the geek, then Drawio is the straight-A student. It can handle almost any diagramming need you can think of. Powerful, free, and feature-rich.

Highlights

Near-commercial features, completely free
Desktop and online versions
Multiple storage options (Google Drive, OneDrive)
Rich template library, plus custom libraries

Frame0 - Hand-drawn style prototyping tool

https://frame0.app

A prototyping tool that is great for engineers to validate ideas quickly. It makes it easy to sketch product concepts. The project started in 2024-09, has released 7 versions, and is still in beta, perfect if you like to try new things.

Highlights

Cross-platform: macOS, Windows, Linux
Smooth interaction
Great for polished presentation diagrams
Real-time collaboration

As an R&D Leader, How to Organize Tech Sharing

Mon, 09 Dec 2024 05:53:26 GMT

Purpose

First, we need to clarify why we’re doing this. What value does it create, and what outcomes do we expect? The “everyone” here includes three parties: the company, the team, and the individual. Each has different expectations from tech learning and sharing. Here are the key purposes:

Learn new technologies, improve ourselves, and build a knowledge system
Build experience in professional domains
Improve communication and presentation skills
Use learning and analysis to advance product architecture, performance, and tooling
Accumulate technical materials and form long-term company value

How to Start

Once the purpose is clear, we need to think about how to make it happen. Adult learning and training are hard to push forward. Heavy delivery workload, human inertia, and whether leaders truly support it are all common blockers, which is why many teams give up halfway. To avoid these pitfalls:

Align management: tech learning and sharing is long-term work, and impact should be evaluated over time
Ensure content quality: content cannot be too shallow
Define a clear cadence and responsibilities for both speakers and audiences
Choose topics based on the audience, focusing on what they care about

How to Do It

Reduce the responsibility denominator. In a 12-person team, the denominator is 12, and people tend to think “it’s not just me,” so they don’t take initiative. Make it small groups instead, with a denominator of 2 or 3. Groups tend to be more reliable than pure self-drive.

Our current R&D team has three roles:

RD: 4
FE: 5
QA: 3

We form three groups and rotate tech learning and sharing in sequence. A group can have multiple speakers per session or just one.

2. Decide the Cadence

The cadence must be explicit so both speakers and listeners can prepare.

Frequency: once every 2 weeks (groups rotate)
Time: Thursday at 19:15
Duration: within 60 minutes

Speaker Preparation

Share the outline: one week in advance (by the previous Thursday)

Audience Preparation

Read the materials: within the week before the session. Studies show that if you don’t pre-read, you may only understand about 20% during the session. Don’t waste your time—read beforehand.

3. Content Selection

The goal of learning and sharing is to raise the team’s technical level and expand its cognitive range. We suggest two types of content: technical topic sharing and free-theme sharing.

How to Choose Content

Selection should vary by team context and product architecture. Plan quarterly, learn the team’s management and technical weaknesses, what people want to improve, and what the product needs next. Then choose topics accordingly (we currently use Lagou Education for training).

Targeted improvement in a specific technical area
Quickly applicable to current architecture and product
Often needs a shallow-to-deep progression

Broaden perspective
Content can be wider, even inviting non-technical colleagues
Requires strong logical expression to explain unfamiliar topics clearly

Alternate between technical topics and free themes.

Reject Simple

If you want the content and format to be great, you need great content. We reject overly simple topics like “how to install X” or “getting started with Y.”

4. Team Atmosphere

Tech learning and sharing improves team culture, strengthens mutual understanding, and increases cohesion.

Provide drinks and snacks, create a relaxed atmosphere, and people will be more willing to join. Initiative improves naturally.

Summary

Beyond learning and sharing, we should also practice gratitude.

Thanks to the company for providing the platform; thanks to the speakers for preparing on their own time; thanks to the audience for listening carefully and asking questions.

With a grateful heart, you’ll see better things and reach new heights by thinking from others’ perspectives.

How to Provide Services Through Cloud Computing

Sun, 24 Nov 2024 09:52:09 GMT

Recently I have been helping our marketing team with structured technical descriptions. While choosing topics, I realized how “AI” and “cloud computing” have become buzzwords. Most people only half understand them, but if you do not mention them you seem out of date. Since we are a SaaS provider, writing a cloud-computing explainer from that angle felt like a good direction.

Definition

Cloud computing is a technology that provides computing resources (such as servers, storage, databases, networks, and software) over the internet. It allows users to access and use these resources on demand without purchasing and maintaining local hardware or software. The core idea is to centralize computing resources and deliver them as services over the network.

Service Models

Cloud computing has three service models, from bottom to top: IaaS -> PaaS -> SaaS.

Deployment Options

There are three common deployment options: public cloud, private cloud, and hybrid cloud.

[Our current deployment] Public Cloud

Provided by cloud vendors (e.g., Alibaba Cloud, Huawei Cloud, AWS, Microsoft Azure)

Pros: cost-effective, scalable, flexible, high availability, and strong security (network and data security)
Cons: relies on third-party providers; for some sensitive data (e.g., national security related), it may not be the best choice

Private Cloud

A cloud environment built for a single organization, hosted on-premise or in a third-party data center, with higher security (after a company and product reach a certain stage, this is typically chosen by key accounts with strict security requirements).

Pros: stronger data security and privacy
Cons: sacrifices some scalability and flexibility; high costs (servers and operations staff)

Hybrid Cloud

If you want the scalability and security of public cloud while keeping data on-premise to satisfy data residency laws, or to support compute closer to customers, hybrid cloud is a good fit.

Pros: combines advantages of public and private cloud, easy to scale resources and storage while still accessing data and apps locally
Cons: complex to manage; requires dedicated VPN or leased lines between public and private clouds; high demands on network speed and stability

IaaS Layer

We primarily use Alibaba Cloud’s infrastructure for networking and storage to build our SaaS foundation.

PaaS Layer

Non-application Services

We mainly use Alibaba Cloud’s development platforms and environments to build and deploy application services, such as container services and serverless computing.

Application Services

Video conferencing capabilities
Document collaboration capabilities
Live streaming capabilities
AIGC
…

SaaS Layer

We provide services to users through the SaaS layer.

Benefits of SaaS

Cost efficiency: users do not need to buy expensive hardware or software licenses; they pay as needed
Fast deployment: users can start by registering an account, reducing the time for traditional install and configuration
Easy to scale: users can add or remove functional modules as needed
Automatic updates and maintenance: the SaaS provider handles updates and maintenance, so users stay on the latest version without worrying about upgrades or security patches
Convenient access: users can access SaaS services from any internet-connected device, supporting remote and mobile work and increasing flexibility
Security: SaaS providers invest heavily in security such as encryption, backups, and disaster recovery, often more reliable than what small and medium businesses can manage themselves

Spending Less: Slimming Down My Subscriptions

Sun, 01 Sep 2024 10:01:31 GMT

As I use more software and services, I realized I get a lot of subscription payment emails every month. I recently reviewed all my internet subscriptions and found plenty of places to save money.

Tencent Video

Category	Cycle and Cost	Need level
Video content	RMB 168/year	Just pick one platform

This used to be my only video subscription, mainly because I was following a bunch of anime adapted from web novels. Tencent Video does have more anime IP than most platforms. The ones I follow:

Soul Land
Battle Through the Heavens
Swallowed Star
Jade Dynasty
Throne of Seal

Xiaomi Video Membership + Bundled iQIYI

Category	Cycle and Cost	Need level
Video content	RMB 392/year	Nice to have

Mainly because we have a Xiaomi TV at home and the kids want to watch.

Apple Music

Category	Cycle and Cost	Need level
Music	RMB 11/month	Nice to have

I got three free months when I bought a device during the pandemic, then kept renewing. Since I am deep in the Apple ecosystem, the experience is smooth. I usually listen to nostalgic songs while commuting or coding at work.

Fan Book (formerly Fan Deng Reading)

Category	Cycle and Cost	Need level
Knowledge subscription	RMB 365/year	Not needed / will not renew

I have listened for four years. In the first two years I listened a lot because there were many books I had not heard of. After two years, I had basically finished most of the content on the platform, so I do not plan to renew. These days my main inputs are physical books and ebooks.

Shengdong Hutong

Category	Cycle and Cost	Need level
Knowledge subscription	RMB 365/year	Not needed / will not renew

I listen to “Shengdong Morning Coffee” every morning for about 15 minutes to catch up on tech news in China and abroad. I love the quality of this team: young, energetic, and professional. Out of pure fandom, I joined their membership. It unlocks some paid podcasts (Not Just Money, Jumping Into the Rabbit Hole), but the ROI is not great. If I am to renew next year, I hope they can give me more reasons. For now, I do not really recommend it or need it.

WeRead (WeChat Reading)

Category	Cycle and Cost	Need level
Knowledge subscription	RMB 19/month	Needed / will renew

This year I used WeRead to replace my old Apple Books + ZLibrary freebie setup.

Two main reasons:

Notes can be managed in one place and exported, which is convenient.
The audiobook experience is great, especially for commuting.

AgentNEO (VPN)

Category	Cycle and Cost	Need level
Network service	RMB 240/year	Needed / will renew

A hard requirement. If you know, you know.

Blog Maintenance Costs

Category	Cycle and Cost	Need level
Domain	RMB 156/year	Needed / will renew

Right now my blog costs are only for the domain (GoDaddy). Everything else is free.

Category	Service	Cost
Code hosting	GitHub (Hugo)	Free
Deployment hosting	Vercel (auto on Git commit)	Free
Image hosting	https://sm.ms	Free (5GB)
CDN	Cloudflare	Free
Domain	GoDaddy	RMB 156/year

Duolingo (Family Plan)

Category	Cycle and Cost	Need level
Learning	RMB 0/year	Needed / will renew

A great tool for language learning. I am using a friend’s family plan for free. You can also find group buys on platforms like Xiaohongshu, which is pretty affordable. Do not buy it alone, it is too expensive.

Cubox-Pro

Category	Cycle and Cost	Need level
Knowledge capture	RMB 98/year	Needed / will renew

My main information inputs come from three sources:

Books (print or digital)
RSS subscriptions
WeChat public accounts
Other social apps

For everything except books, I often read something once and want to go deeper later or save it. I put those items into Cubox. When writing a blog post, planning, or thinking through a topic, I search tags and reprocess that information. For me, Cubox is essential.

iCloud

Category	Cycle and Cost	Need level
Cloud storage	RMB 21/month (200GB)	Not needed / will not renew

iCloud is great for syncing photos and files across devices, but once I exceeded 200GB it stopped being cost-effective. The 200GB plan is RMB 21/month; the next tier is 2TB at RMB 68/month, over RMB 800/year, which hurts. After some research, I chose a NAS solution to solve file sync and backup once and for all, saving that annual 800+ cost. Of course the initial NAS investment is higher than 800, so it is still about long-term cost. I compared ZSpace and UGREEN and finally chose the UGREEN DX4600+ with an 8TB drive. Two months later, UGREEN released a new NAS with software quality issues, so I was lucky that the older software is more stable. The experience has been good so far.

Todoist

Category	Cycle and Cost	Need level
Productivity tool	RMB 30/month	Not needed / will not renew

I am not a heavy task manager. After about six months of subscription, I felt the value was not there for me. I eventually replaced Todoist with a Notion ToDo List template.

Tutu Bookkeeping

Category	Cycle and Cost	Need level
Productivity tool	RMB 8/month	Not needed / will not renew

I am a slightly obsessive programmer, so whether I am coding, writing docs, or dealing with daily information, I like to quantify things and find a sense of accomplishment in visualized data. I started using Tutu Bookkeeping after a recommendation from SSPAI. After about six months, I ran into two big issues: I did not log transactions promptly, and I was not comfortable recording sensitive financial data in an internet product. So for timeliness and privacy, I dropped it.

I switched to the open-source solution beancount. I do bookkeeping quarterly. I export raw transactions from WeChat and Alipay into Excel, do some light processing with a Python script, and generate beancount-formatted text. After minor tweaks, it is ready. The final result looks like this:

Sam’s Club Premium Membership

Category	Cycle and Cost	Need level
Shopping	RMB 680/year	Needed / will renew

In middle age, I have discovered that going to the supermarket can be entertainment. My wife and I both love shopping at Sam’s Club. The quality is top-tier, and in Beijing they now offer 1-hour delivery, which is fantastic. If we order something, it arrives within an hour. The premium membership gives cashback that you can use on next month’s purchases, so once you count the cashback, it is basically the same as the 200+ regular membership.

[Podcast] Shengdong Morning Coffee

[Podcast] Between Reality and Fiction

[AIGC] 302.ai

[RSS] RSS feeds are one of my most important channels for high-quality information. If you need it, add my contact and I will send it.

After optimizing, my annual cost dropped from RMB 3,200 to RMB 2,100 (still expensive). Next I plan to fully leverage my NAS capabilities and maybe replace even more paid subscriptions.

AIGC Application Exploration

Sun, 21 Jan 2024 10:19:36 GMT

Generative AI (Artificial Intelligence Generated Content)

Definition

Baidu Baike

Generative AI (AIGC) is a major sign of the transition from AI 1.0 to AI 2.0.

The accumulated fusion of technologies such as GAN, CLIP, Transformer, Diffusion, pretrained models, multimodal techniques, and generative algorithms has triggered the AIGC explosion. Continuous algorithmic iteration and pretrained models drive a qualitative leap in AIGC capabilities. Multimodality expands the content polygon of AIGC, giving it more general and stronger foundational abilities.

AIGC is a milestone for human society and AI. In the short term it changes basic productivity tools; in the medium term it will reshape social production relations; in the long term it will lead to a qualitative leap in productivity. In this transformation of tools, relations, and productivity, the value of data as a factor of production is greatly amplified.

Core generative models keep making breakthroughs.
Pretrained models triggered a qualitative leap in AIGC capability.
Multimodal techniques expand AIGC content diversity, making it more general.

Changes in productivity, production relations, and means of production

Ecosystem

Foundation Models

Main AIGC pretrained models abroad

Scenarios and Applications

Tool AI
AI toolbox
AI navigation

International AIGC Apps

China AIGC Apps

General Use Case Examples

ChatGPT

ChatPDF

ChatPDF lets you drop in a PDF under 200 pages, and it quickly “reads” the whole document. It can generate a summary of key points, and you can ask follow-up questions. It answers based on the document.

Last weekend I helped my child review. I uploaded a PDF of his math mistakes; the right side shows its answers. The accuracy was pretty good.

Tencent Hunyuan Assistant

Search “Tencent Hunyuan Assistant” in WeChat mini programs.

It integrates multiple scenarios: chat, drawing, coding, etc.

Tongyi Tingyu

Search “Tongyi Tingyu” in WeChat mini programs.

An audio transcription tool that can record speech in real time, translate live, and summarize.

Wenxin Yiyan

In the Wenxin Yiyan app.

It integrates multiple scenarios: chat, drawing, coding, etc.

Runway

https://runwayml.com

hotoke.ai

Hotoke.ai is a ChatGPT-based “Buddha” platform created by a Japanese developer. It has helped millions of people with their worries.

I asked a question people often think about when they are lost: “What is the meaning of life?”

Learning Use Case Examples

Duolingo

A popular free language-learning app that uses gamification. I have used it for a long time and the experience is excellent.

The US version added GPT-4 features: Max.

It offers two main GPT-4-based features: Explain My Answer and Roleplay.

Explain My Answer

The goal is to help students not only know whether their answer is correct or incorrect, but also understand why. Whether their answer is right or wrong, they can talk to the assistant to get a brief explanation and ask for examples or clarification. This helps students understand and fix mistakes and improve learning outcomes.

Roleplay

This feature lets students practice real-world conversations with virtual characters. They can access it by tapping on a character and earn XP for completing challenges. In these challenges, students practice dialogue in different scenarios, such as planning a future vacation with Lin, ordering coffee at a cafe in Paris, buying furniture with Eddy, or inviting friends to go hiking. These exercises help students build language skills across contexts.

Youdao (AI Speaking Tutor)

Similar to Duolingo’s roleplay scenarios, it simulates conversation scenarios, but the experience is average.

API documentation

Mini program available for hands-on experience

Grammarly

Writing assistant that improves writing skills.

Khanmigo

Khanmigo is a personalized learning tool built on GPT-4 and other models. It provides modes like “Tutor Me” and subject-based quizzes. It is most accurate for K-12 topics, but less accurate for niche subjects.

Quizlet (Q-Chat)

Q-Chat is ChatGPT’s AI tutor inside Quizlet. Quizlet is an online learning platform with interactive tools to help students understand and remember knowledge. Q-Chat works like a personalized AI tutor, helping each student based on their learning style and progress. It can have deep, question-driven conversations with students on any topic, using a Socratic approach to spark active learning and deepen understanding.

YouTube demo video: https://www.youtube.com/watch?v=VLUGMzbJ3_M

MathGPT (Not Released)

MathGPT focuses on explaining problems, solving them, and algorithms. It can generate random math problems and provide systematic answers. It supports both Chinese and English, and works on mobile and PC.

Based on current data, MathGPT can cover high school, middle school, and elementary school math. Algebra, word problems, and calculations can all be solved with MathGPT, with clearer reasoning and logic.

Adapt to change, embrace AIGC, learn to use AIGC tools, and combine them with your work and life so AIGC helps us become better.

Conversational assistants: help search, summarize, and organize efficiently (use ChatGPT abroad if possible; otherwise Tencent, Baidu, or Alibaba options)
Image and video assistants: use tools like Midjourney and Runway to accelerate posters and outward-facing copy
Coding assistants: use GitHub Copilot, Bito, and others to speed up coding

My 2023

Sun, 07 Jan 2024 01:44:56 GMT

The end of the year was busy: a major product release, helping my kid study for finals, and so on. I did not wrap up my 2023 until January 7. Here is a recap across life, work, reading, and writing.

Life

The biggest change in 2023 was a new family member: a cute toy poodle named Coffee Bean (our “younger son”). Our family is now me, my wife, our older son (the kid), and our younger son (Coffee Bean). The moment I open the door and both the kid and dog rush at me (sometimes jumping onto me) is the part of every day I look forward to most. It has been a great stress reliever for both work and life.

2023 also featured the most expensive National Day holiday I have ever had: a deep tour of hospitals. At the end of September, my son was diagnosed with mycoplasma pneumonia, which rapidly became severe. We went through a tense search for a hospital bed and were finally admitted to New Century Children’s Hospital. After 12 days of treatment (including two lung lavages), he was discharged safely.

My biggest takeaways:

Healthcare is expensive.
Medical resources are scarce.
After COVID, infectious diseases are so damn intense.
High-end medical insurance is a must (I bought AIA coverage immediately in November).

I did not travel much in 2023, but I did visit Sanya, Guangxi, Changbai Mountain, and Yanji. People say reading ten thousand books is not as good as traveling ten thousand miles. In a time that does not feel so great, do not overwork or over-study. Take time to see the world, learn and understand it from a different perspective. Travel really helps calm the mind and ease anxiety.

In 2023 I wrote five posts, about one every two-plus months. By the end of the year I was honestly ashamed of my laziness. The main reason is that I did not summarize and organize my work and technical learnings enough, so there was not much to share. Next year I will try for one post per month.

A Note to Myself

When I look back, my life is basically either coding or writing plans in front of a screen, or spending time with my kid and walking the dog, with the occasional trip. That is middle age in a nutshell. It may lack passion and big dreams, but overall I am not too anxious (sometimes a little). I do not have strong material desires, no major debt, I work in a field I chose and have built for more than a decade, the company vibe is great, my boss is nice, and I still have time with family. That is pretty good.

In 2024, I hope I can keep reading more, travel more, spend time with family, and continue cultivating myself on the road of life: live honestly, follow my heart, go with the flow, and stay at ease.

AI Coding Assistants

Fri, 21 Jul 2023 01:44:56 GMT

Why Use AI Coding Assistants

Improve coding efficiency and adapt to the new era of programming.

The era of large models is creating “super individuals” who are highly efficient, skilled with tools, and able to deliver business value quickly.
Most tools require a VPN for access; there is no way around it.
Pay attention to information security: do not copy/paste secrets or important code into prompts.

Tool Overviews

Bito (official site: Bito)

Personal plan (free)

Overview

Bito AI is a general-purpose assistant. Developers can ask technical questions, generate code from natural language prompts, and get feedback on existing code. Key features:

Code generation: describe requirements in natural language and Bito generates code in different languages (e.g., Java).
Command syntax: ask about any technical command and its usage (e.g., git).
Test cases: generate test cases for given code.
Explain code: explain selected code snippets and how they work.
Add comments: generate comments for functions or methods.
Improve performance: suggest performance optimizations.
Security checks: check code for known security issues.
Learn concepts: explain technical concepts (e.g., B+ trees).

Installation

Usage

YouTube videos
- https://www.youtube.com/@bitoai/featured
- https://www.youtube.com/watch?v=q42hqwT-jsg
Documentation
- https://docs.bito.ai/getting-started/bito-ai-examples

GitHub Copilot (official site: Copilot)

Personal plan ($10 / month). I can share a referral link (DM me).

Overview

GitHub Copilot provides highly valuable code completion and generation. It can generate lines of code or entire functions based on context. With intelligent completion and generation, Copilot improves development efficiency and lowers the effort of writing code. It further pushes AI adoption in software development so developers can focus more on design and algorithms.

Installation

IDEA plugin
VS Code plugins
- Copilot
- Copilot Labs (chat assistant)

Usage

https://docs.github.com/zh/copilot/quickstart

Codeium (official site: Codeium)

Personal plan (free). Experience is better in VS Code; a bit worse in IDEA.

Overview

Codeium is an AI code generation platform with these key capabilities:

Code generation: describe requirements in natural language and generate code. Supports Java, Python, JavaScript, and more.
Code search: search code snippets by keywords from a large codebase.
Code explanation: explain selected code in natural language.
Code optimization: analyze existing code and suggest improvements in quality or algorithms.
Code editing: built-in editor with autocompletion, linting, formatting, etc.

Installation

Usage

https://codeium.com/chat

Suggested Setup

Personal Experience

Bito

A very powerful assistant. The default templates are handy and make it easy to insert code into chats and ask questions about it. The IDEA and VS Code plugins both feel good.

GitHub Copilot

Excellent at understanding code context (the more you use it, the better it learns you). It can complete code quickly. For routine tools and algorithms, you can often rely on generated code to build.

Codeium

Completion is average. The IDEA plugin is not great, but the VS Code experience is decent. Its features overlap somewhat with Bito.

Using IDEA

Bito + GitHub Copilot (if you have budget)

Using VS Code

Bito + Codeium (if you have budget, switch to GitHub Copilot)

How to Handle Production Issues

Fri, 28 Apr 2023 01:44:56 GMT

What Is a “Production Issue”

In the product lifecycle, a production issue happens after release.

A production issue means the service provided to users is fully or partially unavailable, performs poorly, or the user experience is bad. In early product stages, speed of new features often takes priority over quality to win market share, which accumulates technical debt. When some of that debt explodes, it causes production issues, lowers customer satisfaction, and can even lead to direct financial loss.

Production Issue Flow Overview

Discover -> Handle -> Summarize -> Feedback

Severity

Severity	Description	Handling
Critical	Highest level, system or service fully down or unusable	Take emergency measures immediately
Major	Second level, system or service partially down or restricted	Take measures as soon as possible
Normal	General level, no obvious impact to system or service	Handle within a reasonable time to prevent escalation
Minor	Lowest level, small issues or anomalies with no impact	Address gradually during routine maintenance

Handling Process

User Feedback & Monitoring Alerts

On-call Strategy

Weekly rotation: Every Monday at 10:00 AM, the on-call person for the week is announced in the DingTalk group, as shown below.

Error Analysis Report

Every day we analyze logs and push a log error classification report to the DingTalk alert group.

Before leaving each day, the on-call RD analyzes the report and, if it is a production issue, reports it to QA for tracking.

Postmortem Report

Incident Postmortem (Critical)

Biweekly Review Report (Overall)

How to Understand User Stories

Sun, 05 Mar 2023 01:44:56 GMT

What Is a User Story

A user story is a short and simple description of a feature that delivers value to users or customers, and that a team can deliver within an iteration.

A user story should answer three questions:

Who is it for? -> expected type
What do we build? -> I want
Why do we build it? ->

So the typical format is:

As a , I want , so that .

As a , I want to , so that .

Example: As a registered user, I want to download my photo to my profile so that other users can see what I look like.

How to Write User Stories

Hierarchy

Criteria

According to the INVEST criteria, user stories should be:

Independent
- Not dependent on other stories (in reality this is hard, because some stories truly depend on predecessors).
Negotiable
- Not fixed; can be discussed.
Valuable
- A user story must bring clear, measurable business value. This means technical practices like refactoring, code cleanup, and architecture design are not stories. It also means a story is a vertical slice; it cannot be just a frontend or backend task.
Estimable
- Clear enough that the delivery team can estimate its size.
Small
- Should not exceed what one or two developers can deliver in a single iteration.
Testable
- The story must be verifiable with clear acceptance criteria.

Our User Story Template

Story Writing FAQ

Q1: Can stories be split by different clients?

A1: If a story involves two clients (mobile and PC), and the story points are large and handled by different FE engineers, you can split by client.

Q1: If story points are very large, should we split the story?

A1: The standard for splitting is whether it can deliver “user-visible value.” It is not directly related to the story point size itself.

How to Estimate Story Points

What Is a Story Point

A story point is an estimate of the difficulty and workload of a story, essentially a unit of work. Story points are not a commitment and are not directly tied to actual completion time.

”Inflation” of Story Points

Code quality, developer skill, and familiarity with the business all affect delivery speed. But if different developers estimate story points by completion time, everyone will end up with a different point scale, which leads to “inflation.”

Planning Poker Estimation

Reference: Planning Poker

Planning Poker

Thu, 02 Mar 2023 01:44:56 GMT

What Is Planning Poker

A consensus-based estimation method (a game) used to estimate effort for Scrum iteration tasks. By evaluating together as a team, the variance becomes smaller.

Implementation Steps

Roles

Participant	Description
Scrum Master (QA)	Facilitates the estimation process
PM	Explains and describes the stories
RD / FE	Developers who estimate
RD TL / FE TL	Judges the reasonableness of story points

Round 1

1. PM explains the story
1. RD / FE place numbered cards face down; Scrum Master calls reveal and everyone flips at once (Fibonacci values: 1, 2, 3, 5, 8, 13, 20, 40)
1. Explain estimation differences and discuss
1. Reach consensus

Round 2

1. For stories with differing points, RD / FE explain their estimation differences and reach consensus
1. Scrum Master writes the final story points into the story list
1. Everyone ranks priorities and fills them into the story list

Fibonacci Sequence and Planning Poker

Planning poker uses the Fibonacci sequence as story points. The Fibonacci sequence is a numeric series introduced in the 13th century that explains certain formations in nature, like branching in trees. Each next value is the sum of the previous two: 0, 1/2, 1, 2, 3, 5, 8, 13, 20, etc.

For agile estimation, some values are adjusted, resulting in the following series: 1, 2, 3, 5, 8, 13, 20, 40, 100, as shown below:

Point convention: 1 Story Point ~= 1 person-day in ideal conditions (based on average team capability).

Card	Meaning
0	No effort needed
?	Cannot estimate
Others	Interpret the number as the point value

As an Engineering Leader, How I Manage Iterations

Tue, 07 Feb 2023 01:44:56 GMT

Preface

Iteration management looks different at every company. As product stages, team size, and resources change, the process evolves through practice and optimization. In my team, we review our iteration process every six months and make adjustments.

Below is the iteration workflow we optimized in early 2023.

What Problems Iteration Management Solves

In the diagram above there are three phases. The “product development” phase is the iteration process described below.

Development Process Methodology

Principles and Framework

Note: For the overall iteration cycle, we split it into larger stages that resemble a “waterfall” approach.

For requirement understanding and delivery, we use an agile approach and deliver/accept at the story level.

Principles We Follow

People’s initiative at the core: deliver higher-value products efficiently and with quality, and help team members become the best version of themselves.

Framework We Use

Scrum + Kanban

The goal of Agile is to help us discover how bad things are as early as possible, and start managing the situation early.

Process Overview

Roles

Role	Responsibility
Scrum Master	Agile coach, drives iterations; currently handled by QA
PM	Product manager, produces and explains the PRD
RD	Backend engineer, responsible for backend implementation
FE	Frontend engineer, responsible for frontend implementation
QA	Test engineer, responsible for iteration quality
UI	UI designer, responsible for interaction and visual design

Detailed Breakdown

Requirements Phase

Kick Off Meeting

Goal: A kickoff, as the name suggests. Communicate product value, why we are doing it, and the iteration goals. No deep dive into product or prototype details.

Participants: Everyone

Actions

PM -> present the product value in the PRD
Master -> create the iteration version in the project management platform
Master -> create a communication group in the communication tool

Impact Analysis Meeting

Goal: Understand existing system functionality and analyze the impact of this iteration

Participants: RD, FE, QA

Actions

RD or FE -> explain current system features and analyze impact

PRD Review Meeting

Goal: Discuss requirements, implementation logic, and rule details to reach team-wide alignment on understanding (typically no more than 2 meetings).

Participants: Everyone

Actions

PM -> deliver the final PRD

UI Review Meeting

Goal: Discuss page and interaction implementation to reach a shared understanding of UI/UX (typically no more than 2 meetings).

Participants: Everyone

Actions

UI -> deliver page and interaction designs

Design and Estimation Phase

Plan Meeting

Goal: Confirm iteration scope, including business work, technical work, and backlog or bugs.

Participants: Everyone

Actions

Master -> organize the plan meeting and confirm scope (review backlog and bugs in the project management platform)

Design Review Meeting

Goal: Discuss design, standards, implementation, data migration, historical tech debt, and release plans.

Participants: RD or FE, QA

Actions

(As needed) RD -> produce design doc and review it (see How to write a design doc TODO)
- First review: RD internal design review
- Second review: RD + FE API design review
(As needed) FE -> produce design doc and review it

Estimate Meeting

Goal: Based on the confirmed scope from the plan meeting, estimate effort and assign priorities and owners.

Participants: RD, FE, QA

Actions

Master -> organize the estimate meeting and confirm effort (estimation method: planning poker TODO)

“R&D Cycle” Plan Release

Actions

QA -> publish the iteration “R&D cycle” plan

Master -> update all iteration info in the project management platform and officially enter coding

Test Case Review Meeting

Goal: Review completeness and quality of test cases.

Participants: RD, FE, QA

Actions

QA -> organize the review (test cases maintained in the test case management platform)

Development Phase

Daily Stand-up Meeting

Goal: See TODO: How to run daily stand-ups

Participants: Everyone

Actions

Master -> run the meeting, focusing on dependencies, blockers, risks, etc.

Desk Check and Story Testing

Actions

RD & FE -> initiate story acceptance (how to do story acceptance TODO). After it passes, deploy to FAT for testing; QA records results.
QA -> test the story; if there are no functional issues, close the story testing

Mid-term Check

Goal: Surface issues and anticipate risks

Actions

Master -> at the midpoint of the iteration, organize a progress review to detect delay risks

Test and Acceptance Phase

System Testing

Actions

RD & FE -> prepare deployment docs and submit deployment tickets
QA -> in FAT
- validate deployment docs
- smoke testing
- test cases (black-box, white-box)
- regression testing (black-box, white-box)

Acceptance

Actions

QA -> in UAT
- validate deployment docs
PM -> product acceptance
UI -> UI acceptance
QA -> in UAT
- regression testing (black-box, white-box)

Release Phase

Release

Actions

QA -> submit deployment ticket for approval, then release
QA -> fill out release notes (internal)

Iteration Retrospective Phase

Review Meeting

Goal: Summarize iteration process and quality issues, run a retrospective, and optimize process, code quality, and collaboration.

Participants: Everyone

Actions

Master -> gather iteration data and write the report
Master -> organize a full-team review and discussion

Tools and Platforms

	Tool or Platform
Project management platform	https://www.teambition.com/
Communication tool	DingTalk
Test case management platform	https://metersphere.io

As an R&D Leader, How to Quantify Iteration Quality

Wed, 16 Feb 2022 04:01:57 GMT

Why Quantify

Let Facts Speak

Using data to reflect collaboration outcomes and the quality of delivered results is a way to surface problems from another objective angle. Avoid subjective judgments like “I think” or “I feel” when evaluating complex problems.

Diverse Roles, Complex Collaboration

An iteration often involves many roles. Their actions and outputs all influence “iteration quality,” and those changes are usually complex. That’s why we should collect objective facts across multiple dimensions and evaluate iteration quality comprehensively.

Iteration Should Also Iterate

Product optimization and iteration are continuous. At different stages, companies and products have very different quality requirements. The famous project management triangle is a trade-off among time, cost, and scope, and quality is affected accordingly. As the company and product evolve, we need to collect data from multiple angles and continuously analyze iteration quality, so we can make the right trade-offs for the current context and reach our product goals.

How to Quantify

Within our team, we use a standard to evaluate iteration quality, called the “Iteration Quality Evaluation Standard,” composed as follows.

Components of the Iteration Quality Evaluation Standard

4 dimensions + 1 score

4 dimensions
- PM requirement changes (30%)
- User Story delivery delays (10%)
- User Story smoke test pass rate (20%)
- Bugs (40%)
1 score
- Iteration quality score

Full score 10, pass 6	Dimension	Weight
PM requirement changes (after PRD finalization)	30%	3
User Story delivery testing delays	10%	1
User Story smoke test pass rate	20%	2
Bugs (n = total bugs / total story points)	40%	4

Example

Above is a recent set of iteration data from my team. Context:

Story points: 1 story point = 1 person-day
Iteration flow: develop & test by story → overall regression testing → PM & UI acceptance → release → retrospective

What Datafication Brings to the Team

The team is accountable for the iteration
- By presenting iteration quality with data, the team can review problems rationally and find solutions. The shared goal is: make the iteration better.
Make “problems” transparent
Provide a solid data foundation to improve team iteration effectiveness (any change can be reflected in data)

Things We Want to Do but Haven’t Yet

By connecting the agile tools we use and automating iteration data collection, we can generate an “iteration quality score” and compare historical data across dimensions, thereby improving efficiency.

Agile Practice for a Small Team 2.0

Wed, 18 Aug 2021 05:47:18 GMT

Phase 1.0

1.0

Everything Starts from Requirements

During the 1.0 phase, we found many practical issues in implementation and usage. So we improved and upgraded the overall agile process to adapt to changes in team iterations.

Requirement Source Categories

Because requirement channels are diverse, we categorized sources for easier management:

Requirement Code	Description	Iteration Version Code
Feature	Product-value driven self-initiated iteration	feature/sprintXX, e.g. feature/sprint40
CS	User-side requirements from Customer Success managers	cs/yyyyMMdd, e.g. cs/20210713
Tech	Tech improvement or refactoring initiated by R&D	tech/module-or-refactor-name, e.g. tech/mtms, tech/res
Hotfix	Online issue fixes triggered by incident flow	hotfix/yyyyMMdd, e.g. hotfix/20210713

Currently, our team only triggers Sprints for feature requirements. Other types do not trigger Sprints.

Iteration Process

Overall, the iteration has 6 stages:

PRD Review: Product requirement review, user story review
Estimate: High-level design, effort estimation (planning poker)
Sprint Start: Create Sprint in Jira, version number, register user stories, create Gantt
Sprint In Progress: Iteration in progress, daily standup at 10am to sync
Deploy: PM & UI acceptance, release
Sprint Review: Iteration retrospective

1. PRD Review

PRD Review Flow

Participants

Lead: PM
Participants: RD FE QA UI

Input

PRD for this iteration. (Path: “PM Team” → “Product Requirements” → “2021 Product Requirements” → “Product Self-Iteration” → “PRD page of a specific iteration”)

Output

Final prototype discussion in Lanhu
Final user story discussion (Path: “DingTalk Drive” → “3-R&D Dept” → “User Story Breakdown” → “SprintXX” → “User story files in the folder”)

PRD

User stories in DingTalk

2. Estimate

We use Planning Poker. It’s a consensus-based estimation method (a game) used to estimate development effort in Scrum iterations. Through team consensus, deviation becomes relatively small.

What Is Planning Poker

Planning Poker uses the Fibonacci sequence as Story Points. The Fibonacci sequence is a series of numbers introduced in the 13th century, used to explain patterns in nature (e.g., tree branching). The next value is the sum of the previous two: 0, 1/2, 1, 2, 3, 5, 8, 13, 20. For agile estimation, some numbers are adjusted to: 1, 2, 3, 5, 8, 13, 20, 40, 100.

Poker legend (point convention: 1 Story Point = 1 person/day)

Card	Explanation
0	No work needed
Infinity	Large task
?	Cannot estimate
A cup of coffee	Less than 0.5
Others	Interpret points literally

Estimation Flow

Participants

Lead: PM
Participants: RD FE QA

Input

User stories

Output

Estimates and priorities for user stories

Steps (usually consensus is reached within 2 rounds)

PM describes the user story

RD / FE team members play cards face down (Fibonacci values: 1, 2, 3, 5, 8, 13, 20, 40)

Reveal cards simultaneously

Explain estimation differences and discuss

Reach estimation consensus

Estimation record

After estimation, we fill the results into the story list in DingTalk Drive and confirm story priorities.

3. Sprint Start

Sprint Creation & Start Flow

Participants

Lead: Scrum Master (QA)
Participants: RD FE QA

Input

User stories & estimation results

Output

Sprint & Gantt

Create Epic

Create Story

Sprint board

Gantt

4. Sprint In Progress

Daily Flow

Participants

Lead: Scrum Master (QA)
Participants: RD FE QA

Input

Standup sync

Output

User story updates

5. Deploy

Acceptance & Release Flow

Participants

Lead: Scrum Master (QA)
Participants: PM UI RD FE

Input

UAT environment

Output

Iteration release
Sprint Done

6. Sprint Review

Review Flow

Participants

Lead: QA
Participants: PM UI RD FE

Input

Test report

Output

Retrospective results

Test report

Summary

The above is the current 2.0 version of our R&D team’s agile iteration process. The key improvements ahead are introducing CI, automated testing, and other approaches to further improve testing efficiency and surface issues faster.