Token导航 LogoToken导航

Agent Skill · 研究检索

deep-work(深度工作)

deep-work 用于查找、检索和筛选相关信息,适合在 Codex、Claude、Cursor、Gemini CLI 中需要根据关键词、任务场景或来源线索快速定位候选结果时使用。可结合来源仓库、安装命令和原始 README 继续核验具体用法。安装前建议确认权限范围、维护状态,以及是否会触发联网、命令执行或文件读写。

deep-work 是研究检索类 Agent Skill,可用宿主为 Codex、Claude、Cursor、Gemini CLI,通过GitHub 仓库获取,权限为只读;许可证信息还不明确,第三方扫描已通过,按安装量在研究检索类的 73,675 个 Skill 里排第 53,927。

总安装 519周安装 21GitHub Stars 4下载 163核验 2026-05-01

安装说明

本站只整理中文说明,不托管安装包,也不代用户安装。
GitHub
许可证
unknown

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:deep-work(深度工作)
来源仓库:https://github.com/petekp/agent-skills
仓库路径:skills/deep-work
安装命令:
npx skills add https://github.com/petekp/agent-skills --skill deep-work
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。该命令会通过 npx skills 从第三方来源获取 Skill;本站只展示命令,不托管安装包,也不自动执行。

skills.shnpx skills
npx skills add https://github.com/petekp/agent-skills --skill deep-work

这个 Skill 是做什么的

研究检索类 Skill 主要解决「找资料、核对事实、把来源整理清楚」的问题,通常会让助手先检索再总结。

deep-work 采用结构化工作流分离思考与编码阶段,强制先制定并审核书面计划再实施。适用于复杂多步骤任务分解、代码重构或需严格前期调研的开发场景,强调可追溯性。通过研究→计划→标注→待办→实现→反馈六阶段循环,所有工件持久化保存于 .claude/ 目录。

使用前需确认项目根目录可写,并注意其会创建本地标记文件记录各阶段进展与决策依据。deep-work 属于研究检索类 Skill,可作为该场景下的辅助能力补充。

来源仓库内路径:skills/deep-work

上手步骤

4 步
  1. 1

    确认自己的宿主工具

    公开资料标注支持 Codex、Claude、Cursor、Gemini CLI;先确认自己用的是其中之一,再按对应工具的 Skills 目录放置。

  2. 2

    拿到安装命令

    可以直接复制的命令是 npx skills add https://github.com/petekp/agent-skills --skill deep-work;命令里的占位符换成自己的路径或密钥再执行。

  3. 3

    安装后重启宿主

    命令执行完要完全退出宿主工具再打开,让新装的 Skill 被重新扫描到,否则列表里可能看不到。

  4. 4

    按最小权限试用

    它标注为只读,先用只读任务试一次,确认行为符合预期再扩大使用范围。

常见问题

6 条

deep-work 怎么安装?

在终端执行 npx skills add https://github.com/petekp/agent-skills --skill deep-work。装完完全退出宿主工具再打开,让它重新扫描本地 Skill。

deep-work 支持哪些 AI 工具?

公开资料标注支持 Codex、Claude、Cursor、Gemini CLI。其他工具只要能读取本地 Skill 目录,一般也能手动放进去,但要自己确认触发方式是否一致。

deep-work 需要什么权限?

权限级别是只读。该 Skill 主要提供规则、说明或参考内容,本身偏只读;真正读写文件、联网或执行命令仍取决于宿主 Agent 的任务。

deep-work 是免费的吗?

公开资料没有标注明确许可证,收费和商用限制都无法确认。正式使用前请到来源仓库核对授权说明。

deep-work 用的人多吗?

公开数据是累计安装 519 次,近一周安装 21 次,下载 163 次,GitHub 4 星。这些数字来自第三方目录统计,只能当作热度参考,不代表质量或安全性。

deep-work 安全吗?

第三方扫描结果为通过,但这只覆盖已知的静态检查项,不能替代自己读一遍源码和安装命令。

值得选的地方和要先想清楚的地方

值得选的地方

  • 同时适配 4 种工具,换客户端不用重新找替代品。
  • 第三方扫描已通过,常见的高风险行为没有命中。
  • 标注为只读权限,试用阶段不容易误改本地文件。

先想清楚的地方

  • 许可证没有明确标注,能不能商用、能不能再分发都不确定。
  • 第三方目录整理,版本更新和行为细节以来源仓库为准。

适不适合你

更适合

  • 已经在用 Codex,想直接扩展它能力的人。
  • 需要处理研究检索类任务、又不想每次都手写提示词的人。
  • 只想先试试效果、不希望它改动本地文件的人。

不太适合

  • 需要明确授权才能用到公司项目里的场景。
  • 需要厂商级技术支持和 SLA 的生产系统。

研究检索类 Skill 对比

6 个同类
Skill总安装GitHub 星权限许可证扫描
find-skills

寻找技能 · Codex、Claude

31,199,81316,486操作浏览器MIT有提醒
self-improving-agent

自我改进 Agent · OpenClaw

16,464,0023,396接触敏感数据MIT-0通过
soultrace

灵魂追踪 · Codex、Claude

7,588,94412需要联网未明确有提醒
self-improving

自我完善 · OpenClaw

6,829,2841,046需要联网MIT-0通过
microsoft-foundry

微软代工厂 · Codex、Claude

6,614,144764调用外部服务未明确有提醒
deep-work

深度工作 · 本页

5194只读未明确通过

按安装量取同分类前几名;数字来自第三方目录统计,只作热度参考,不代表质量结论。

SKILL.md 原文

Deep Work

Structured workflow that separates thinking from typing. Never write code until a written plan has been reviewed and approved.

Workflow Overview

Research → Plan → Annotate (repeat 1-6x) → Todo List → Implement → Feedback

All artifacts persist as markdown files in .claude/ (not the project root).

Phase 1: Research

Deeply read the relevant codebase before doing anything else. Write findings to .claude/research.md.

How to research:

  1. Read every file in the target area — not just signatures, but implementations, edge cases, error handling
  2. Trace data flows end-to-end
  3. Identify existing patterns, conventions, utilities, and caching layers
  4. Note integration points with other parts of the system
  5. Write a detailed report to .claude/research.md

Research depth signals: Read deeply. Understand intricacies. Go through everything. Surface-level reading is not acceptable. Continue until you have a thorough understanding.

Template: See references/research-template.md for the research document structure.

Critical rule: Stop after writing research.md. Do not proceed to planning until the user has reviewed the research and confirmed it's accurate.

Phase 2: Plan

Write a detailed implementation plan to .claude/plan.md. Base the plan on the actual codebase — read source files before suggesting changes.

Plan contents:

  • Explanation of the approach and rationale
  • Code snippets showing actual changes (not pseudocode)
  • File paths that will be modified or created
  • Considerations and trade-offs
  • Things explicitly not being changed and why

Template: See references/plan-template.md for the plan document structure.

Reference implementations: If the user provides reference code from other projects, study it and adapt the approach to fit the current codebase's patterns.

Critical rule: End the plan with "Ready for your review. Add inline notes directly to .claude/plan.md and tell me when to address them." Do not implement.

Phase 3: Annotation Cycle

The user adds inline notes directly into plan.md. When they say "address my notes" or similar:

  1. Read .claude/plan.md thoroughly, finding all user annotations
  2. User annotations are any text that wasn't in your original plan — look for corrections, questions, overrides, and constraints
  3. Address every single annotation: update sections, remove rejected approaches, incorporate domain knowledge
  4. Remove the annotation text itself after addressing it (keep the plan clean)
  5. End with: "All notes addressed. Review again or say 'add the todo list' when satisfied."

Critical rule: Do not implement. The phrase "don't implement yet" is a hard constraint. Repeat the annotation cycle until the user explicitly approves.

Phase 4: Todo List

When the user approves the plan, add a granular task checklist to .claude/plan.md:

## Tasks

### Phase 1: [Phase Name]
- [ ] Task 1 — specific, actionable description
- [ ] Task 2 — specific, actionable description

### Phase 2: [Phase Name]
- [ ] Task 3 — specific, actionable description
- [ ] Task 4 — specific, actionable description

Each task should be small enough to complete in one focused step. Include all phases needed to fully implement the plan.

Critical rule: Do not implement yet. Wait for user confirmation to begin.

Phase 5: Implementation

When the user says "implement" or "go":

  1. Execute every task in the plan sequentially
  2. After completing each task, update .claude/plan.md — change - [] to - [x]
  3. Do not stop until all tasks are completed
  4. Run the project's type checker / linter after each phase to catch issues early
  5. Do not add unnecessary comments, docstrings, or type workarounds (any, unknown)

Code quality rules during implementation:

  • No unnecessary comments or jsdocs
  • No any or unknown types (TypeScript projects)
  • Run typecheck/lint continuously
  • Follow existing codebase patterns exactly

Phase 6: Feedback

During implementation, the user may provide terse corrections. These are sufficient because full context exists in the plan and session:

  • "You missed the dedup function" → implement what was missed
  • "Move this to the admin app" → relocate as directed
  • "Wider" / "2px gap" / "still cropped" → adjust and re-check
  • "Make it look like the users table" → read that reference, match it exactly

On reverts: If the user says "I reverted everything" — re-read the current file state, narrow scope to exactly what they specify, and re-implement cleanly.

Prompt Reference

See references/prompts.md for ready-to-use prompts for each phase that the user can copy and adapt.

平台分布

Codex

37.14%
按下载量换算61

Claude

26.93%
按下载量换算44

Cursor

18.78%
按下载量换算31

Gemini CLI

8.26%
按下载量换算13

安全审计

Gen Agent Trust Hub

通过

Socket

通过

Snyk

通过

权限和风险

只读

该 Skill 主要提供规则、说明或参考内容,本身偏只读;真正读写文件、联网或执行命令仍取决于宿主 Agent 的任务。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。目前只有一个公开来源,建议同时核对源仓库。

继续浏览同类 Skills