real-work

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Real Work

实际工作

Turn planning into a durable, resumable artifact. The plan file — not the conversation — is the source of truth: it records what to do, what's done, how it was verified, and how to deploy. Any future agent can resume from it with zero prior context.

Use when planning multi-step / multi-session work that may outlive the current session. Skip for trivial single-session tasks.

将规划转化为可持久化、可恢复的工件。计划文件——而非对话——是事实来源：它记录了要做什么、已完成什么、如何验证以及如何部署。任何后续Agent都可以从零上下文开始继续执行。

适用场景：规划可能超出当前会话时长的多步骤/多会话工作。琐碎的单会话任务无需使用。

1. Reach complete understanding first

1. 先达成完整共识

Do not write the plan until scope is fully understood. Relentlessly ask the user questions until you both share a complete understanding with no gaps — treat an unasked question as a future bug.

Don't stop at the first round; keep going until no ambiguity, assumption, or open decision remains. Probe edges: scope boundaries (in/out), dependencies, constraints, success criteria, data, environments, deployment, failure cases.
Surface every assumption for the user to confirm. If an answer opens a new unknown, ask the follow-up — drill down recursively.
Use
```
AskUserQuestion
```
for concrete choices. When done, summarize the full scope back and only proceed once the user confirms nothing is missing.

在完全理解工作范围之前，不要编写计划。持续向用户提问，直到双方对工作范围达成无遗漏的完整共识——未提出的问题视为未来的隐患。

不要停留在第一轮提问；持续提问，直到没有歧义、假设或未决事项。探究边界：范围界限（包含/排除）、依赖关系、约束条件、成功标准、数据、环境、部署方式、故障场景。
列出所有假设供用户确认。如果某个答案引出新的未知问题，继续跟进提问——递归式深入探究。
对于具体选项，使用
```
AskUserQuestion
```
工具。完成提问后，总结完整的工作范围并反馈给用户，仅在用户确认无遗漏后再继续。

2. Write the plan

2. 编写计划

Save to

plans/<descriptive-name>.md

in the repository root (create

plans/

if needed). Use this self-documenting template:

markdown

undefined

保存到仓库根目录的

plans/<描述性名称>.md

文件中（若

plans/

目录不存在则创建）。使用以下自文档化模板：

markdown

undefined

<Work Title>

<工作标题>

<1-2 sentence goal and scope.>

<1-2句话说明目标和范围。>

For Future Agents

给后续Agent的说明

As work proceeds: mark checkboxes

- [x]

as items complete; when a phase is done, set its status to

Complete

and write its Phase Summary (what was done, key decisions, anything needed to continue with zero context); run the phase's Verification Plan and record the result before moving on. When all phases are done, fill in Final Recap and Deployment Plan.

工作推进过程中：将复选框

- [x]

标记为已完成项；当一个阶段完成时，将其状态设置为

Complete

并撰写阶段总结（已完成工作、关键决策、无需上下文即可继续工作所需的所有信息）；执行该阶段的验证计划并记录结果，之后再进入下一阶段。所有阶段完成后，填写最终回顾和部署计划。

Phase 1: <Title>

阶段1：<标题>

Status: Not started

<concrete, actionable item>
<concrete, actionable item>

状态：未开始

<具体、可执行的任务项>
<具体、可执行的任务项>

Verification Plan

验证计划

<command/check the agent can run autonomously, with expected result>

<Agent可自主运行的命令/检查，以及预期结果>

Phase Summary

阶段总结

(write when phase completes)

（阶段完成后撰写）

Phase 2: <Title>

阶段2：<标题>

Status: Not started

<actionable item>

状态：未开始

<可执行任务项>

Verification Plan

验证计划

<autonomous check>

<自主检查项>

Phase Summary

阶段总结

(write when phase completes)

（阶段完成后撰写）

Final Recap

最终回顾

(write when all phases complete: summary of the entire piece of work)

（所有阶段完成后撰写：整个工作的总结）

Deployment Plan

部署计划

(write when all phases complete: step-by-step deployment instructions)

undefined

（所有阶段完成后撰写：分步部署说明）

undefined

Common mistakes

常见错误

Vague items — each checkbox is a concrete task ("Add retry logic to
```
PaymentClient.Charge
```
"), not a theme ("improve payments").
Non-autonomous verification — give runnable commands with expected output, not "test it manually".
Wrong location — always the repo-root
```
plans/
```
folder.
Pre-filling summaries — phase summaries, recap, and deployment plan stay as placeholders until that work actually completes.

任务项模糊——每个复选框对应一个具体任务（例如“为
```
PaymentClient.Charge
```
添加重试逻辑”），而非一个主题（例如“优化支付功能”）。
非自主验证——提供可运行的命令及预期输出，而非“手动测试”。
存储位置错误——务必存放在仓库根目录的
```
plans/
```
文件夹中。
提前填写总结——阶段总结、最终回顾和部署计划在对应工作实际完成前保持占位符状态。