Token导航 LogoToken导航

Agent Skill · 研究检索

do-work(做工作)

do-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。

do-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,权限为权限未标注;许可证信息还不明确,第三方扫描有未通过项,按安装量在研究检索类的 73,675 个 Skill 里排第 62,370。

总安装 336周安装 14GitHub Stars 公开资料未说明下载 112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:do-work(做工作)
来源仓库:https://github.com/knews2019/skill-do-work
仓库路径:skills/do-work
安装命令:
npx skills add https://github.com/knews2019/skill-do-work --skill do-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/knews2019/skill-do-work --skill do-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。

do-work 用于查找、检索和筛选相关信息。适合在 Codex、Claude、Cursor、Gemini CLI 中根据关键词、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。

do-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/do-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/knews2019/skill-do-work --skill do-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

    权限没有明确标注,第一次使用时只给一个无关紧要的任务,观察它到底读了哪些文件和网络。

常见问题

6 条

do-work 怎么安装?

在终端执行 npx skills add https://github.com/knews2019/skill-do-work --skill do-work。装完完全退出宿主工具再打开,让它重新扫描本地 Skill。

do-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。

do-work 需要什么权限?

权限级别没有明确标注。第一次使用时建议只给无关紧要的任务,观察它实际读写了哪些内容。

do-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。

do-work 用的人多吗?

公开数据是累计安装 336 次,近一周安装 14 次,下载 112 次。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。

do-work 安全吗?

第三方扫描有未通过项,风险较高。除非你确认过源码,否则不建议直接安装到常用环境。

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。

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。
  • 第三方扫描存在提醒项或未通过项,需要人工逐条确认。
  • 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。

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
do-work

做工作 · 本页

336—权限未标注未明确未通过

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

Do-Work Skill

A unified entry point for task capture and processing.

Actions:

  • capture requests: Capture new tasks/requests → creates UR folder (verbatim input) + REQ files (queue items), always paired
  • verify requests: Evaluate captured REQs against original input → quality check
  • work: Process pending requests → executes the queue
  • clarify questions: Batch-review Open Questions from completed work → user answers, confirms, or skips
  • review work: Post-work review → requirements check, code review, acceptance testing, and testing suggestions
  • present work: Client-facing deliverables → briefs, architecture diagrams, value propositions, Remotion videos
  • cleanup: Consolidate archive → moves loose REQs into UR folders, closes completed URs
  • commit: Commit uncommitted files → analyzes, groups atomically, traces to REQs
Core concept: The capture requests action always produces both a UR folder (preserving the original input) and REQ files (the queue items). Each REQ links back to its UR via user_request frontmatter. This pairing is mandatory for all requests — simple or complex.
Capture ≠ Execute. The capture requests action captures requests. The work action executes them. These are strictly separate operations. After capture finishes writing files and reporting back, STOP. Do not start processing the queue, do not begin implementation, do not "helpfully" transition into the work action. The user decides when to execute — always. The only exception is if the user explicitly says something like "add this and then run it" or "capture this and start working" in the same invocation.
Human time has two optimal windows. The system is designed to maximize the value of human attention: 1. Capture phase (capture requests action) — The user is present, actively thinking about the request. This is the best time for back-and-forth: clarifying ambiguities, resolving contradictions, making scope decisions. Use AskUserQuestion with concrete choices here. Every question must present options — never open-ended "what do you mean?" prompts. 2. Batch question review (clarify questions action) — After the build phase completes everything it can without feedback, any remaining pending-answers REQs are surfaced as a batch. The user reviews all builder-decided questions together, confirms or adjusts, and resolved REQs re-enter the queue. Between these windows, the build phase runs autonomously. Builders never block on Open Questions — they mark them - [~] with best-judgment reasoning and create pending-answers follow-ups when they return via do work clarify.

Routing Decision

Step 1: Parse the Input

Examine what follows "do work":

Check these patterns in order — first match wins:

PriorityPatternExampleRoute
1Empty or bare invocationdo work→ help menu
2Action verbs onlydo work run, do work go, do work start→ work
3Verify keywordsdo work verify, do work verify requests, do work check, do work evaluate→ verify requests
4Clarify keywordsdo work clarify, do work questions, do work pending→ clarify questions
5Review keywordsdo work review, do work review work, do work review code, do work code review, do work audit code→ review work
6Present keywordsdo work present, do work present work, do work showcase, do work deliver→ present work
7Cleanup keywordsdo work cleanup, do work tidy, do work consolidate→ cleanup
8Commit keywordsdo work commit, do work commit changes, do work save work→ commit
9Version keywordsdo work version, do work update, do work check for updates→ version
10Changelog keywordsdo work changelog, do work release notes, do work what's new, do work what's changed, do work updates, do work history→ version
11Descriptive contentdo work add dark mode, do work [meeting notes], do work capture request [the request]→ capture requests

Step 2: Preserve Payload

Critical rule: Never lose the user's content.

Single-word rule: A single word is either a known keyword or ambiguous — it is never "descriptive content."

  • Matches a keyword in the routing table (e.g., "version", "verify", "cleanup") → route to that action directly.
  • Doesn't match any keyword (e.g., "refactor", "optimize") → ambiguous. Ask: "Do you want to add '{word}' as a new request, or did you mean something else?"

Only route to capture requests when the input is clearly descriptive — multiple words, a sentence, a feature request, etc.

If routing is genuinely unclear AND multi-word content was provided:

  • Default to capture requests (adding a task)
  • Hold onto $ARGUMENTS
  • If truly ambiguous, ask: "Add this as a request, or start the work loop?"
  • User replies with just "add" or "work" → proceed with original content

Action Verbs (→ Work)

These signal "process the queue": run, go, start, begin, work, process, execute, build, continue, resume

Clarify Verbs (→ Clarify Questions)

These signal "review pending questions": clarify, answers, questions, pending, pending answers, blocked, what's blocked, what needs answers

Note: This routes to the work action with mode: clarify — see work.md "Clarify Questions" section.

Verify Verbs (→ Verify Requests)

These signal "check request quality": verify, verify requests, check, evaluate, review requests, review reqs, audit

Note: "check" routes to verify requests ONLY when used alone or with a target (e.g., "do work check UR-003"). When followed by descriptive content it routes to capture requests (e.g., "do work check if the button works" → capture requests).

Note: "audit" alone routes to verify requests. "audit code" and "audit implementation" route to review work (see Review Verbs below).

Review Verbs (→ Review Work)

These signal "review the completed work": review, review work, review code, code review, audit code, audit implementation, review REQ-NNN, review UR-NNN

Note: "review requests" and "review reqs" route to verify requests (priority 4), not review work. "review" alone or followed by a target/code-related word routes to review work (priority 5). The review work action also runs automatically as part of the work pipeline — see work.md Step 7.

Present Verbs (→ Present Work)

These signal "generate client-facing deliverables": present, present work, showcase, deliver, pitch, client brief

Note: do work present (no target) presents the most recent completed UR. do work present all or do work present portfolio enters portfolio mode. do work present UR-003 or do work present REQ-005 targets specific work.

Cleanup Verbs (→ Cleanup)

These signal "consolidate the archive": cleanup, clean up, tidy, consolidate, organize archive, fix archive

Commit Verbs (→ Commit)

These signal "commit uncommitted files atomically": commit, commit changes, commit files, save changes, save work

Changelog Verbs (→ Version)

These signal "show release notes": changelog, release notes, what's new, what's changed, updates, history

Note: "updates" (plural) routes to changelog display. "update" (singular) routes to update check. Both are handled by the version action.

Content Signals (→ Capture Requests)

These signal "add a new task":

  • Descriptive text beyond a single verb
  • Feature requests, bug reports, ideas
  • Screenshots or context
  • "add", "create", "I need", "we should"

Examples

Help Menu (bare invocation)

When invoked with no arguments (do work), show a help menu with available actions and example prompts:

do-work — task queue for agentic coding tools

  Capture requests:
    do work add dark mode to settings
    do work the search is slow and the header is misaligned
    do work [paste meeting notes, specs, or a screenshot]

  Process the queue:
    do work run

  Verify & review:
    do work verify requests     Check capture quality against original input
    do work review work         Review completed work (requirements + code + acceptance)

  Present to client:
    do work present work        Generate client brief, architecture, video, and interactive HTML explainer
    do work present all         Portfolio summary of all completed work

  Other actions:
    do work clarify             Answer pending questions from completed work
    do work cleanup             Consolidate the archive
    do work commit              Analyze and commit uncommitted files atomically
    do work version             Check version / updates
    do work changelog           Show release notes

Do not ask "Start the work loop?" — just print the help menu and wait.

Routes to Work

  • do work run → Starts work action immediately
  • do work go → Starts work action immediately

Routes to Clarify Questions

  • do work clarify → Presents all pending-answers REQs for batch review
  • do work questions → Same as clarify
  • do work answers → Same as clarify
  • do work pending → Same as clarify
  • do work what's blocked → Same as clarify

Routes to Verify Requests

  • do work verify requests → Evaluates most recent UR's REQs
  • do work verify → Evaluates most recent UR's REQs
  • do work verify UR-003 → Evaluates specific UR
  • do work check REQ-018 → Evaluates the UR that REQ-018 belongs to
  • do work evaluate → Evaluates most recent UR's REQs
  • do work review requests → Evaluates most recent UR's REQs

Routes to Review Work

  • do work review work → Reviews the most recently completed REQ
  • do work review → Reviews the most recently completed REQ
  • do work review REQ-005 → Reviews a specific completed REQ
  • do work review UR-003 → Reviews all completed REQs under that UR
  • do work code review → Reviews the most recently completed REQ
  • do work review code → Reviews the most recently completed REQ

Routes to Present Work

  • do work present work → Generates deliverables for most recently completed UR
  • do work present → Same as present work
  • do work present UR-003 → Generates deliverables for specific UR
  • do work present REQ-005 → Generates deliverables for specific REQ
  • do work present all → Portfolio summary of all completed work
  • do work present portfolio → Same as present all
  • do work showcase → Same as present work

Routes to Cleanup

  • do work cleanup → Consolidates archive, closes completed URs
  • do work tidy → Same as cleanup
  • do work consolidate → Same as cleanup

Routes to Commit

  • do work commit → Analyzes and commits all uncommitted files atomically
  • do work commit changes → Same as commit
  • do work save work → Same as commit

Routes to Changelog (via Version)

  • do work changelog → Displays changelog (newest at bottom)
  • do work release notes → Same as changelog
  • do work what's new → Same as changelog
  • do work updates → Same as changelog
  • do work history → Same as changelog

Routes to Capture Requests

  • do work add dark mode → Creates REQ file + UR folder
  • do work the button is broken → Creates REQ file + UR folder
  • do work [400 words] → Creates REQ files + UR folder with full verbatim input

Payload Preservation Rules

When clarification is needed but content was provided:

  1. Do not lose $ARGUMENTS - keep the full payload in context
  2. Ask a simple question: "Add this as a request, or start the work loop?"
  3. Accept minimal replies: User says just "add" or "work"
  4. Proceed with original content: Apply the chosen action to the stored arguments
  5. Never ask the user to re-paste content

This enables a two-phase commit pattern:

  1. Capture intent payload
  2. Confirm action

Action Dispatch

Each action has an action file with full instructions. How you execute it depends on your environment's capabilities.

ActionAction fileContext to pass
capture requests./actions/capture.mdFull user input text
work./actions/work.md(none needed)
clarify questions./actions/work.mdmode: clarify
verify requests./actions/verify-requests.mdTarget UR/REQ or "most recent"
review work./actions/review-work.mdTarget REQ/UR or "most recent"
present work./actions/present-work.mdTarget REQ/UR, "most recent", or "all"
cleanup./actions/cleanup.md(none needed)
commit./actions/commit.md(none needed)
version./actions/version.md$ARGUMENTS

If subagents are available

Dispatch each action to a subagent. The subagent reads the action file and executes it — the main thread only sees the routing decision and the returned summary.

  • work and cleanup: Run in the background if your environment supports it. Print a status line (e.g., "Work queue processing in background...") and return control to the user immediately.
  • capture requests, clarify questions, verify requests, review work, present work, version: Run in the foreground (blocking). These need user interaction or produce small immediate output.
  • Screenshots (capture requests only): Subagents can't see images from the main conversation. Before dispatching, save screenshots to do-work/user-requests/.pending-assets/screenshot-{n}.png, write a text description of each, and include the paths + descriptions in the subagent prompt.

If subagents are not available

Read the action file directly and follow its instructions in the current session. The action files are designed to work as standalone prompts — no subagent infrastructure required.

On failure

Report the error to the user. Do not retry automatically.

Suggest Next Steps

After every action completes, suggest the next logical prompts the user might want to run. Use fully qualified action names so the user can copy-paste directly.

After capture requests:

Next steps:
  do work verify requests     Check capture quality before building
  do work run                 Start processing the queue

After work (queue processing):

Next steps:
  do work review work         Review the completed work
  do work present work        Generate client-facing deliverables
  do work clarify             Answer any pending questions

After verify requests:

Next steps:
  do work run                 Start processing the queue
  do work [describe changes]  Capture additional requests

After review work:

Next steps:
  do work present work        Generate client-facing deliverables
  do work run                 Process follow-up REQs (if any were created)
  do work [describe changes]  Capture new requests

After present work:

Next steps:
  do work present all         Generate portfolio summary (if multiple URs completed)
  do work [describe changes]  Capture new requests

After commit:

Next steps:
  do work review work         Review the committed changes
  do work [describe changes]  Capture new requests

After clarify questions:

Next steps:
  do work run                 Process answered questions
  do work clarify             Continue answering (if skipped any)

Rules:

  • Only suggest prompts that provide value given the current state (e.g., don't suggest do work run if the queue is empty)
  • Use the full action name (verify requests, not just verify; review work, not just review)
  • Keep it to 2-3 suggestions max — don't overwhelm
  • Format as a simple list the user can scan and copy

平台分布

Codex

38.91%
按下载量换算44

Claude

30%
按下载量换算34

Cursor

20.17%
按下载量换算23

Gemini CLI

8.79%
按下载量换算10

安全审计

Gen Agent Trust Hub

通过

Socket

可疑

Snyk

未通过

权限和风险

权限未标注

当前来源未能明确判断权限范围,默认进入异常复核队列。

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。来源安全扫描存在提醒或未通过项,安装前请自行复核。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