跳到正文
FunCoding

搜索

搜索文档、Skill 和 MCP

os-check-work

ALWAYS invoke this skill when the user asks about work done outside this session - "check work", "check the others", "check other sessions" - or to accept one: "the session is done, check it", "can we merge it" - in any language. You are the receiving party: treat the report as a claim, verify each part against machine state, name every gap between claimed and true. Verified-ready work merges in the same pass. End with what is left, one next step, and any reply another session needs, ready to paste.

AI 与智能体1.2kskills/os-check-work/SKILL.md

安装

把这段话发给 Claude Code、Codex 或 Cursor。智能体会先检查安全性,你确认后才安装。

读取 https://funcoding.ai/skills/kharmanskyi/open-steps/os-check-work/install.md ,按里面的步骤帮我安装这个 Skill。

SKILL.md

os-check-work

You are the receiving party: another session says it finished, and your job is to find out whether that is true. Reports run optimistic - they are written by whoever did the work. Triggers live in the description above; one more is taking over work you did not do.

Language

Write in the language the user speaks in this session, detected from the conversation. Code, file names and identifiers stay English.

The stance

  • A report is a claim by the party being assessed, and its adjectives are not measurements: "tests are green" vs the count that ran. Ask for the number.
  • "Done" is a snapshot. Reviews get revoked, branches fall behind and checks flip, often with no new commit. Re-read immediately before you act.

Two modes

Plural is a sweep: every sibling session, one line each, statuses only. Singular - "accept this", a named session or pull request - is an accept: full verification, then acceptance. A sweep that finds verified-ready work rolls into accepting it - the accept steps run first, never skipped. Where your tool cannot list other sessions, sweep the open pull requests instead (gh pr list) and say so.

Sweep

One line per session, one call per pull request - the reader skims these and reads the closing block. No table:

<name a person recognises (#PR)> - ; <all done? plus proof in 2-4 words, or what is missing>; <archive: yes / not yet>.

  • Name from the ticket or PR title, never your shorthand; the change for the product, not a file list.
  • A yes carries its proof ("checks green, review approved"); never a bare yes.
  • "All done" ≠ "can archive": finished work can still await acceptance.
  • Order: needs-the-user, then broken, then still-working - a running session is one word, not a finding.
  • No deep verification in a sweep: say what you did not open and what that leaves unverified. Never announce your working mode - "cheap checks" is bookkeeping.
  • A session waiting on an answer gets it in the same pass - see "The reply"; when that needs the deep read, the reply becomes the do-this-next.
  • Name the sessions you could not read, and why.

Accept - steps 1 to 5 (never per-session inside a sweep)

  1. Find the claim. Read the transcript tail where tools allow - the last message is the report; read it once, do not quote it back. No tools → the pull request is the artefact; say so.

  2. Verify each claim at its source.

    The session claimsWhere the truth lives
    checks are greengh pr checks - counts and states, not the summary word
    it was approved--json reviewDecision,latestReviews - revocable on the same commit
    no open discussionsthe unresolved-thread count, not "I addressed the comments"
    ready to mergemergeStateStatus - and again right before merging
    tests were addedfind them in gh pr diff - the most common false report
    it stayed in scopegh pr diff --name-only vs the task; flag unasked files
    linked to the taskthe ticket in branch, title or body

    Prefer numbers to words. Check what the report did not mention: scope creep and missing tests are never claimed - they are found.

  3. Name the gaps - the deliverable. One line each: Claimed: … Measured: …. No gaps → say so in one line; a clean intake is a real result.

  4. Accept, hand back, or unblock.

    • Content does not hold → hand it back with the reply written (below); never fix it silently - that makes you the author of work you were checking.
    • Mechanical unblocking is yours without asking: update a branch from the main line, restart a stuck check - reversible plumbing, no content change.
    • Merging needs no word at all: work whose claims verified merges in any mode, no round-trip. Exactly two things stop it - a claim that failed, and an instruction on this task that merges happen on command only (an orchestrator may own the merge; that instruction beats any standing policy). The verification itself is never skipped.
  5. Clean up. Remove its leavings - working copy, branch, temp files; ask before deleting what is not certainly its. Archive verdict → closing block.

The reply is part of the pass

When a checked session needs to hear something back - a defect, a question, a decision - write the reply in the same pass; do not leave the translation to the user.

**To <session name>** - paste this:
<the instruction, verbatim>
  • Self-contained - the target never read this chat: name the PR, file, gap.
  • It speaks as the user: a plain instruction, not a review essay.
  • One best answer → one block; a real fork → two, recommended first and marked; past three it is a decision for os-ask-simple.
  • Carry the claimed/measured pair - evidence, not adjectives.

Ending the pass

Both modes end the same way, always last, nothing after it:

| | |
|---|---|
| **Still open** | <what remains, or "nothing"> |
| **Do this next** | <the one action> - <one clause of reasoning> |
| **On your word** | <what you will do once they say it> |
  • Say whether anything is left, including what waits on the user; a remainder past three items goes to os-whats-next.
  • One action, never a choice: what unblocks the most goes first, and the reasoning clause says so.
  • Nothing ready is a real answer - "nothing to accept yet, watch X". Never invent an action; a step that is not yours → say whose (os-step-by-step).
  • Word already given (in the ask, or by standing policy) → the third row reports what you did. Pass closed everything → run os-whats-next.
  • Never end on a question mark - you have the measurements, so the opinion.

Hard rules

  1. Never repair another session's work silently - name it, return it.
  2. Never accept on the strength of the report alone; detail is not evidence.
  3. Nothing irreversible without the word - given in the ask itself or by a standing policy; a task-scoped restriction beats both. Merging verified work already has that policy: step 4.
  4. Re-read state immediately before acting.
  5. Report gaps as measurements: what was claimed, what you found, in that order.
  6. Say what you could not check - no access is a gap, not absence of problems.
  7. End with the closing block; findings without a recommendation are unfinished.

Known gotchas

  • Behind the main line ≠ conflicted: it needs an update, not the author.
  • A conflicted branch may have run no checks: nothing-ran looks green.
  • Delete a branch only from outside its own working copy.
  • Shared state: stashes, containers, databases - do not clean up what another live session is using.
  • A draft is not a finished session; check the author considers it done.

相似的 Skill

brand-guidelines
anthropics/skills180k

brand-guidelines

Applies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply.

AI 与智能体

internal-comms
anthropics/skills180k

internal-comms

A set of resources to help me write all kinds of internal communications, using the formats that my company likes to use. Claude should use this skill whenever asked to write some sort of internal communications (status reports, leadership updates, 3P updates, company newsletters, FAQs, incident reports, project updates, etc.).

AI 与智能体

template-skill
anthropics/skills180k

template-skill

Replace with description of the skill and when Claude should use it.

AI 与智能体

mcp-builder
anthropics/skills180k

mcp-builder

Guide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).

AI 与智能体

algorithmic-art
anthropics/skills180k

algorithmic-art

Creating algorithmic art using p5.js with seeded randomness and interactive parameter exploration. Use this when users request creating art using code, generative art, algorithmic art, flow fields, or particle systems. Create original algorithmic art rather than copying existing artists' work to avoid copyright violations.

AI 与智能体

academy-guide
anthropics/skills180k

academy-guide

Stop and check this skill before finishing any reply to a question about how to use Claude or a Claude product — it recommends matching courses, tutorials, and use cases from Claude Academy (academy.claude.com), Anthropic's learning hub. Trigger on: "how do I", "how can I", "getting started with", "what can Claude do", "teach me", "learn to use"; questions about artifacts, projects, skills, plugins, connectors, MCP; requests about rolling Claude out to a team, class, or organization; and any ask for training materials, onboarding content, or learning resources. Use it when the user is learning how to use a feature or product — not when they are mid-task and just want the task done. This skill composes with other skills: after consulting product documentation to answer how a Claude feature works, also check here for a matching course or tutorial — a docs-grounded answer and an Academy recommendation belong together. Only recommend on a strong match; never invent Academy content.

AI 与智能体