deepworkplan

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

DeepWorkPlan — Methodology Skill (Router)

DeepWorkPlan —— 方法论技能（路由层）

Models matter; context matters more. The DeepWorkPlan skill turns any repository into a structured environment — context, guardrails, and a durable plan — where any coding agent executes reliably on short or long work. It makes the repository "AI-first" —

AGENTS.md

docs/

+ per-module docs +

.agents/

(with the

.claude → .agents

and

.cursor → .agents

symlinks) — and runs structured Deep Work Plans: Lite plans for bounded work and Full task-file plans for longer work; an AI agent creates, refines, executes task-by-task, and resumes. All plan output lands in a gitignored

.dwp/

directory at the repo root (

.dwp/plans/

Source of truth: https://deepworkplan.com. License: MIT.

模型很重要，但上下文更关键。DeepWorkPlan 技能可将任意代码仓库转化为具备上下文、防护规则和可持续计划的结构化环境，让任何编码Agent都能可靠地执行短期或长期工作。它让代码仓库实现「AI优先」—— 包含

AGENTS.md

docs/

+ 各模块文档 +

.agents/

（通过

.claude → .agents

和

.cursor → .agents

符号链接实现）—— 并运行结构化的深度工作计划（Deep Work Plans）：轻量计划用于边界清晰的工作，完整任务文件计划用于长期工作；由AI Agent创建、优化、逐任务执行并支持恢复。所有计划输出都存放在仓库根目录下被git忽略的

.dwp/

目录中（具体路径为

.dwp/plans/

）。

官方参考：https://deepworkplan.com。许可证：MIT。

Start here (first run)

从这里开始（首次运行）

This skill is a self-sufficient entry point: whether a developer arrives from https://deepworkplan.com/init.md or simply installs this skill, the setup plan is the same — and it lives here, so no network is required.

If the repository is not yet AI-first — there is no root

AGENTS.md

and no

.agents/

directory — the recommended first action is to onboard it, even if the developer's request was vague ("set this up", "make this repo AI-first", or a plain install). Before routing anywhere else:

Read the standard locally. Read
```
spec/
```
(five RFC-2119 documents) and
```
shared/adaptation.md
```
. The overriding rule is REASON, do not copy-paste: this skill is the reusable engine; what you produce must be adapted to this repository, never templated.
Run onboarding. Read
```
onboard/SKILL.md
```
and execute it. It is non-destructive: detect existing
```
AGENTS.md
```
,
```
docs/
```
,
```
.agents/
```
, or
```
CLAUDE.md
```
, reconcile rather than overwrite, and ask the developer before replacing anything. The result:
```
AGENTS.md
```
+
```
CLAUDE.md
```
symlink, a reasoned
```
docs/
```
tree, per-module docs, a
```
.agents/
```
kit, and a gitignored
```
.dwp/
```
— the repository becomes the agent harness.
Verify conformance. Read
```
verify/SKILL.md
```
and run it to confirm, objectively, that the repository now meets the standard (AGENTS.md with real commands, the
```
.agents/
```
catalog, the gitignored
```
.dwp/
```
, and so on).
Then plan and execute. With the harness in place, create and execute Deep Work Plans (below) — long-horizon, gated, resumable work an agent can run autonomously for hours.

If the repository is already AI-first but its harness predates this skill —

AGENTS.md

.agents/

exist, yet there is no

DWP standard:

provenance line, the provenance line is older than the standard shipped here, or

docs/TESTING_GUIDE.md

lacks the scoped-invocation and mapping content the standard requires — offer the targeted harness upgrade before routing: reinstalling the skill is the whole upgrade path, so no separate installer exists. Say in one line what is out of date, then run

onboard/SKILL.md

upgrade

mode (Phase 0), which reconciles only the missing or outdated pieces, leaves every handwritten section, custom skill and in-flight plan untouched, and changes nothing on a second run. An in-flight plan keeps its recorded lifecycle; migrating one is a separate, explicit

refine migrate

. If the developer declines, route by intent as normal.

If the repository is already AI-first and current, skip onboarding and route by intent silently — do not announce the detection or the routing decision (no "the repo is already AI-first" / "routing to the create sub-skill" preamble). Just begin the matched sub-skill's flow directly.

该技能是一个自包含的入口：无论开发者是从 https://deepworkplan.com/init.md 跳转而来，还是直接安装本技能，设置流程都是一致的——且流程内置在本地，无需联网。

如果代码仓库尚未实现AI优先——即根目录没有

AGENTS.md

且不存在

.agents/

目录——推荐的首个操作是执行仓库接入，即便开发者的需求很模糊（比如「设置一下」「让这个仓库AI优先」，或是单纯的安装操作）。在路由到其他任何功能之前：

本地读取标准规范。读取
```
spec/
```
（5份RFC-2119文档）和
```
shared/adaptation.md
```
。最高原则是主动推理适配，切勿直接复制粘贴：本技能是可复用的引擎，你生成的内容必须适配当前仓库，绝不能直接套用模板。
执行接入流程。读取
```
onboard/SKILL.md
```
并执行。该流程是非破坏性的：会检测已有的
```
AGENTS.md
```
、
```
docs/
```
、
```
.agents/
```
或
```
CLAUDE.md
```
，采用协调合并而非覆盖的方式，替换任何内容前都会询问开发者。最终产出：
```
AGENTS.md
```
+
```
CLAUDE.md
```
符号链接、结构合理的
```
docs/
```
目录树、各模块文档、
```
.agents/
```
工具包，以及被git忽略的
```
.dwp/
```
目录——代码仓库将成为Agent的运行载体。
验证合规性。读取
```
verify/SKILL.md
```
并运行，客观确认仓库现已符合标准（包含带实际可用命令的AGENTS.md、
```
.agents/
```
目录、被git忽略的
```
.dwp/
```
目录等）。
随后规划并执行。有了运行载体后，即可创建并执行深度工作计划（见下文）——这类工作周期长、有阶段校验、可中断恢复，Agent能够自主运行数小时。

如果代码仓库已经实现AI优先，但适配套件版本早于本技能——即存在

AGENTS.md

.agents/

，但没有

DWP standard:

来源标识行，或来源标识行的版本早于本技能内置的标准版本，或是

docs/TESTING_GUIDE.md

缺少标准要求的范围调用和映射内容——请在路由前提供定向适配套件升级：重新安装本技能就是完整的升级路径，无需单独的安装程序。先用一句话说明哪些内容已过时，然后以

upgrade

模式运行

onboard/SKILL.md

（阶段0），该模式仅合并缺失或过时的部分，不会改动任何手写内容、自定义技能和进行中的计划，且二次运行不会产生任何变更。进行中的计划会保留其已记录的生命周期；迁移计划是一个单独的显式操作

refine migrate

。如果开发者拒绝升级，则按正常流程根据意图路由。

如果代码仓库已经实现AI优先且版本为最新，跳过接入流程，静默根据意图路由——无需告知检测结果或路由决策（不要先说「该仓库已实现AI优先」/「正在路由到创建子技能」这类开场白），直接启动匹配的子技能流程即可。

What it does

功能说明

This is the router. It does not run any flow itself — it maps the developer's intent to the right sub-skill and tells the agent to read that sub-skill's

SKILL.md

and execute it there.

这是路由层。它本身不执行任何流程——只是将开发者的意图匹配到对应的子技能，指示Agent读取该子技能的

SKILL.md

并在对应流程中执行。

Trust boundary (write scope)

信任边界（写入范围）

allowed-tools

includes write-capable

Edit

Write

, and

Bash

because the sub-skills this router delegates to need them; the router itself only reads (the repo tree, the local

spec/

shared/

, and sub-skill files). Writes happen inside the delegated sub-skill, each of which declares its own trust boundary:

Onboard writes
```
AGENTS.md
```
,
```
docs/
```
, per-module docs,
```
.agents/
```
, and appends to
```
.gitignore
```
— reconciling with, never clobbering, what exists, and asking before replacing anything.
Create / refine write plan artifacts under the gitignored
```
.dwp/
```
directory only.
Execute / resume write task outputs, progress, and per-task commits — gated by each task's validation, never committing secrets, never pushing without the developer's instruction.
Addons install or configure anything only after the developer explicitly accepts the offer, always via pinned, verified install paths.

It MUST NOT: make network calls in the core flow, read or commit credentials, run installers unprompted, or write anywhere outside the surfaces above. The full guarantees and a runnable self-audit live in

TRUST.md

allowed-tools

包含具备写入权限的

Edit

、

Write

和

Bash

，因为本路由层委派的子技能需要这些工具；路由层本身仅执行读取操作（读取仓库目录树、本地

spec/

、

shared/

目录及子技能文件）。写入操作都在被委派的子技能中执行，每个子技能都声明了自身的信任边界：

Onboard 写入
```
AGENTS.md
```
、
```
docs/
```
、各模块文档、
```
.agents/
```
，并追加内容到
```
.gitignore
```
——会与已有内容协调合并，绝不会覆盖现有内容，替换任何内容前都会询问开发者。
Create / refine 仅在被git忽略的
```
.dwp/
```
目录下写入计划产物。
Execute / resume 写入任务输出、进度及逐任务提交——受每个任务的校验规则约束，绝不会提交密钥，未经开发者指示绝不会执行推送操作。
Addons 仅在开发者明确接受提议后才安装或配置内容，且始终通过固定的、经过验证的安装路径执行。

严禁以下行为：在核心流程中发起网络调用、读取或提交凭证、未经提示运行安装程序、或写入上述范围之外的任何位置。完整的保障说明及可运行的自审计流程见

TRUST.md

。

For the agent — routing rules

给Agent的说明 —— 路由规则

When the developer wants to plan, execute, or manage structured work, or make a repo AI-agent-ready, match the intent below and read that sub-skill's
SKILL.md
to execute it. Do not answer directly — each sub-skill carries the full step-by-step flow.

Developer says…	Route to
"create a plan", "plan this small fix", "organize this work", "new deep work plan", "/dwp-create"	Create → read `create/SKILL.md`
"execute the plan", "run the plan", "/dwp-execute"	Execute → read `execute/SKILL.md`
"modify the plan", "change the scope", "/dwp-refine"	Refine → read `refine/SKILL.md`
"resume", "continue the interrupted plan", "/dwp-resume"	Resume → read `resume/SKILL.md`
"plan status", "what's left", "/dwp-status"	Status → read `status/SKILL.md`
"verify", "is this repo AI-first?", "check conformance", "/dwp-verify"	Verify → read `verify/SKILL.md`
"make this repo AI-first", "onboard this repo", "set up AGENTS.md + docs + .agents"	Onboard → read `onboard/SKILL.md`
"create/update a skill or agent", "evolve the kit", "/skill-create", "/agent-create"	Author → read `author/SKILL.md`

If the intent is ambiguous between planning and managing existing work, ask the developer which they mean before routing.

当开发者需要规划、执行或管理结构化工作，或是让代码仓库适配AI Agent时，请匹配以下意图，并读取对应子技能的
SKILL.md
来执行。不要直接回答——每个子技能都包含完整的分步执行流程。

开发者的需求示例	路由到
「创建一个计划」「规划这个小修复」「整理这项工作」「新建深度工作计划」、 `/dwp-create`	Create → 读取 `create/SKILL.md`
「执行计划」「运行计划」、 `/dwp-execute`	Execute → 读取 `execute/SKILL.md`
「修改计划」「调整范围」、 `/dwp-refine`	Refine → 读取 `refine/SKILL.md`
「恢复」「继续中断的计划」、 `/dwp-resume`	Resume → 读取 `resume/SKILL.md`
「计划状态」「还剩什么」、 `/dwp-status`	Status → 读取 `status/SKILL.md`
「验证」「这个仓库是AI优先的吗？」「检查合规性」、 `/dwp-verify`	Verify → 读取 `verify/SKILL.md`
「让这个仓库AI优先」「接入这个仓库」「设置AGENTS.md + docs + .agents」	Onboard → 读取 `onboard/SKILL.md`
「创建/更新技能或Agent」「升级工具包」、 `/skill-create` 、 `/agent-create`	Author → 读取 `author/SKILL.md`

如果意图在「规划新工作」和「管理现有工作」之间存在歧义，请先询问开发者的需求，再进行路由。

Normative specification (ships with the skill)

规范性规范（随技能内置）

The methodology's authoritative standard lives at

spec/

— five RFC-2119 documents (

DOCUMENTATION_STANDARD

DWP_SPECIFICATION

AGENT_PROTOCOL

ARCHETYPES

ADDONS

). It ships inside the skill so an agent reads the standard locally — no network needed. The

onboard

flow and

shared/adaptation.md

reference it as the standard to produce. The public, rendered version lives at https://deepworkplan.com/spec.

该方法论的权威标准存放在

spec/

目录下——包含5份遵循RFC-2119规范的文档（

DOCUMENTATION_STANDARD

、

DWP_SPECIFICATION

、

AGENT_PROTOCOL

、

ARCHETYPES

、

ADDONS

）。标准随技能内置，因此Agent可以本地读取标准，无需联网。

onboard

流程和

shared/adaptation.md

均以此作为生成内容的标准。公开的渲染版标准可访问 https://deepworkplan.com/spec 查看。

Shared resources used by every sub-skill

所有子技能共用的资源

```
shared/context.sh
```
— detect repo root, branch, and agent tool; resolve the
```
.dwp/
```
output location.
```
shared/dwp-paths.md
```
— the
```
.dwp/plans/
```
+
```
.dwp/plans/
```
output convention and how to override it.
```
shared/adaptation.md
```
— the reasoning-over-copy-paste principle and the two repository archetypes (individual repo vs orchestrator hub).
```
shared/troubleshooting.md
```
— read only when something is already wrong: discovery failure, stale installation, missing test command, unsupported host capability, inconsistent state.

```
shared/context.sh
```
—— 检测仓库根目录、分支和Agent工具；解析
```
.dwp/
```
输出位置。
```
shared/dwp-paths.md
```
——
```
.dwp/plans/
```
输出目录约定及覆盖方法。
```
shared/adaptation.md
```
—— 「推理优先而非复制粘贴」原则，以及两种仓库原型（独立仓库 vs 编排中心）。
```
shared/troubleshooting.md
```
—— 仅在出现问题时读取：包括检测失败、安装过期、测试命令缺失、主机能力不支持、状态不一致等问题的排查。

Addons — one required local review, four opt-in

附加组件 —— 1个必选本地审查，4个可选接入

The

addons/

area holds the capabilities the

onboard

flow layers onto a repo. The AI Diff Reviewer local review is part of the baseline since standard 2.3.0 (installed by

onboard

Phase 7a, run by every Final Review; its CI surface stays optional). The other four addons are opt-in and never part of the AI-first baseline — a repo is fully conformant with zero optional addons. The first optional addon is devcontainer support.

addons/

目录存放

onboard

流程可叠加到仓库中的功能。AI Diff Reviewer 本地审查自标准2.3.0版本起成为基线的一部分（由

onboard

阶段7a安装，每次最终审查都会运行；其CI功能仍为可选）。其余四个附加组件均为可选接入，不属于AI优先基线的组成部分——即便不安装任何可选附加组件，仓库也完全符合合规要求。第一个可选附加组件是devcontainer支持。