full-cycle

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Full cycle

全流程

Carry work from an idea to a PR a stranger can approve. Spend tokens freely: fan out subagents in dynamic workflows wherever breadth, independent opinions or adversarial checks make the answer better. The human spends attention: they steer and correct, the agent does the deep work.

Five words carry this skill. Tier list: every survey ends in a ranked list with one plain reason per row. Recommended: every question to the human carries the agent's own pick. Receipts: every number was really measured; what was not run is said. Skeptic: every finding is reproduced by a second agent before anyone acts on it. Dry: review ends only when rounds stop producing real findings.

将工作从想法推进到一位陌生评审者也能批准的 PR。可以放心投入 token：在广度、独立意见或对抗性检查有助于改善结果的地方，通过动态工作流并行调度子 Agent。人类把注意力留给引导和纠正，Agent 负责深度工作。

这个 skill 由五个词贯穿。 分级清单（Tier list）：每次调研最终都要给出一份排序清单，每行附一个直白理由。 推荐项（Recommended）：向人类提出的每个问题，都要附上 Agent 自己的首选方案。 实测凭据（Receipts）：每个数字都确实经过测量；未执行的内容必须明确说明。 质疑验证（Skeptic）：采取行动前，每项发现都要由第二个 Agent 复现。 审查收敛（Dry）：只有当审查轮次不再产生实质性发现时，审查才会结束。

Asking the human

向人类询问

Research until you could make the call yourself, then hand the human the call. Each ask lays out what they need to decide and nothing more: the options, what each costs and gains, what it touches. Mark your pick

(recommended)

with one plain reason. The human confirms, picks another, or redirects.

The human switches context all day, and a thread may span weeks. Each message stands alone: answer first, plain words, one picture where a picture is faster than a paragraph. A Mermaid map for structure and flow, a motion schematic for what changes over time (

pr-issue

skill,

motion.md

), a screenshot or slowed real recording for what a user sees. Style is owned by the

agent-to-human

skill.

先充分调研，直到你能够自行做出判断，然后再请人类拍板。每次 ask 只说明人类需要决定的内容：有哪些选项、各自的成本与收益、会影响什么。将你的首选标记为

(recommended)

，并附一个直白理由。人类可以确认、另选方案或重新定向。

人类会全天切换上下文，而一个线程可能持续数周。每条消息都应独立成立：先给出答案，使用平实语言；如果一张图比一段话更快，就只配一张图。结构与流程使用 Mermaid 图，时间上的变化使用动态示意图（

pr-issue

skill、

motion.md

），用户可见内容使用截图或放慢的真实录屏。表达风格由

agent-to-human

skill 负责。

Peers

对标产品

From what the product is, name its peers: open-source apps and libraries that face the same problem. For a coding-agent client that means other agent UIs, editors and CLIs. Read their code, not their marketing. Every research stage answers "how did others solve this problem?" with file paths as evidence.

根据产品形态找出对标产品：解决同一问题的开源应用和库。对于 coding-agent 客户端，就是其他 Agent UI、编辑器和 CLI。阅读它们的代码，而不是营销宣传。每个调研阶段都要回答 “其他人是如何解决这个问题的？”，并以文件路径作为证据。

Stages

阶段

Gates marked ask wait for the human. Everything else runs unattended.

Survey. A workflow fans out across the problem space — for "optimize the app": rendering, startup, bundle, network, memory, server latency. Each agent measures or reads, not guesses, and checks the peers. Return a tier list of areas. Ask which to take.
Deep dive. Inside the chosen area, build or reuse a repeatable measurement, take a baseline, then return a tier list of concrete targets with their measured cost. A target whose cost is an estimate gets a probe: a throwaway fix, measured, reverted. Probes routinely refute the obvious suspect; rank by probe results.
Options. For the top target, 2–4 fixes, always including the answer to "is there a stupid simple solution we're not thinking about?" and the peers' approach. Prototype the serious ones in isolated worktrees and measure them interleaved. For each: upside, downside, what else it touches, and a picture of how the system changes. Ask which to build.
Pre-review. Before coding, one agent attacks the chosen design with "what have we missed?" and returns the final brief.
Implement in its own worktree and branch, following the repo's
```
AGENTS.md
```
/
```
CLAUDE.md
```
and owning specs.
Review until dry. See below.
Evidence. Rebuild before measuring; stale builds lie. Before and after run interleaved, five or more rounds, median with spread. Render or call counts beat milliseconds when timings are noisy. Run the full test suite once. Record media: slow a real recording until the difference is visible, and add a capture-only on-screen marker (a frame counter, a timestamp) when the change is about time.
PR. Write it with the
```
pr-issue
```
skill: before/after table, the options rejected and why, future work, visuals. Slice into one-problem PRs, stacked when they build on each other. Show the human the local media paths before any PR body changes. Ask before opening anything on a repo the human does not own.

标记为 ask 的关卡需要等待人类，其余环节均自动执行。

调研。 工作流在问题空间内并行展开——以“优化这个应用”为例，涵盖渲染、启动、bundle、网络、内存和服务端延迟。每个 Agent 都要进行测量或阅读，而不是猜测，并检查对标产品。返回一份领域的分级清单。询问要选择哪一项。
深入分析。 在选定领域内，构建或复用可重复的测量方式，取得基线，然后返回一份具体目标的分级清单，并附上实测成本。若某目标的成本仅为估算，则进行一次 探针（probe）：实施临时修复、进行测量，再回退。探针经常会推翻最明显的嫌疑对象；根据探针结果排序。
方案。 针对首要目标提出 2–4 个修复方案，其中必须包含对 “我们是不是漏掉了一个简单到愚蠢的解决方案？” 的回答，以及对标产品采用的做法。在隔离的 worktree 中为值得认真实施的候选方案制作原型，并交错测量。逐一说明：收益、代价、还会影响什么，以及系统变化的可视化图示。询问要构建哪个方案。
预审。 编码前，让一名 Agent 用 “我们漏掉了什么？” 攻击选定设计，并给出最终开发简报。
实现。 在自己的 worktree 和 branch 中实现，遵循仓库的
```
AGENTS.md
```
/
```
CLAUDE.md
```
以及相关规范。
反复审查直至收敛。 详见下文。
证据。 测量前先重新构建；过期的构建结果会误导人。变更前后的测量应交错进行至少五轮，报告中位数及波动范围。当计时噪声较大时，渲染次数或调用次数比毫秒数更可靠。完整测试套件运行一次。录制媒体证据：放慢真实录屏，直到差异清晰可见；当改动与时间有关时，在画面中加入仅用于录制的屏幕标记（帧计数器、时间戳）。
PR。 使用
```
pr-issue
```
skill 编写：前后对比表、被否决的方案及原因、未来工作、可视化材料。拆分为每个 PR 只解决一个问题的形式；相互依赖时使用 stacked PR。在修改 PR 正文之前，先向人类展示本地媒体路径。对于人类不拥有的仓库，在开启任何 PR 前都要询问。

Review until dry

反复审查直至收敛

A review is a loop, not an agent.

Find. Parallel reviewers, one lens each:
- correctness: "what have we missed?", proven with throwaway tests;
- "accessibility, other platforms, weird edge cases?": screen readers and keyboard, every runtime the product ships on, phones, slow hardware, and user settings — zoom, locale, RTL, reduced motion, OS themes;
- "can it be better, more performant, simpler?": subtract layers, state and dependencies that earn nothing;
- repo hygiene and upstream readiness.
Sweep. Mechanical unused-code check on everything the change adds:
```
knip
```
,
```
tsc --noUnusedLocals --noUnusedParameters
```
, and a grep proving every new export and returned field has a reader. Reviewers miss what tools see.
Validate. A skeptic per lens reproduces each finding, defaults to false when it cannot, and drops anything that already happens on the base branch.
Fix the confirmed ones; name the rest as follow-ups.
Repeat. Done when two consecutive rounds yield no real finding — only false positives or "looks fine".
Picture check, last round: compare the change against the schematics and videos agreed at the options stage. Behaviour drifted from the picture means fix the code or redraw the picture, and tell the human which.

审查是一个循环，而不是一个 Agent。

发现。 并行审查，每名审查者关注一个视角：
- 正确性：“我们漏掉了什么？”，并用临时测试证明；
- “无障碍、其他平台、奇怪的边界情况？”：屏幕阅读器和键盘操作、产品发布支持的每个 runtime、手机、低端硬件和用户设置——缩放、locale、RTL、reduced motion、OS 主题；
- “它能否更好、更快、更简单？”：删掉没有价值的层次、状态和依赖；
- 仓库整洁度与上游提交就绪度。
清扫。 对变更新增的全部内容执行机械式未使用代码检查：
```
knip
```
、
```
tsc --noUnusedLocals --noUnusedParameters
```
，并用
```
grep
```
证明每个新增 export 和返回字段都有使用方。审查者会漏掉工具能发现的问题。
验证。 每个视角安排一名质疑者复现每项发现；无法复现时默认判为
```
false
```
，并剔除
```
base branch
```
上已经存在的问题。
修复已确认的问题；其余列为后续事项。
重复。 连续两轮都没有产生实质性发现时才算完成——只剩误报或“看起来没问题”。
图示校验，放在最后一轮：将变更与方案阶段确认的示意图和视频进行比较。实际行为与图示发生偏离时，要么修复代码，要么重绘图示，并告知人类采用了哪种处理。

Done when

完成标准

Every ask carried a
```
(recommended)
```
pick and the human's answer is recorded.
The PR shows measured before/after numbers with spread, and names what was not run.
Two review rounds in a row came back dry, and the picture check passed.
Every gate, test suite and build the repo defines passes on the final tree.
A reader with no context understands the change from the PR's first paragraph and its visuals.

每次 ask 都附有一个
```
(recommended)
```
首选方案，并记录了人类的回答。
PR 展示了实测的前后数据及波动范围，并明确列出未运行的内容。
连续两轮审查都没有发现实质问题，且图示校验通过。
仓库定义的所有关卡、测试套件和构建都在最终代码树上通过。
不了解上下文的读者也能从 PR 的第一段和可视化材料中理解改动。