用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/y3owk1n/nix-system-config-v2 --skill review命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Add a skill source or local skill to the Nix configuration.
Scan codebase for deepening opportunities — refactors that turn shallow modules into deep ones.
Create, derive, or update the PRD — the single source of truth every other skill reads.
正在显示 SKILL.md
| name | review |
| description | Two-axis review of a diff: Standards (repo conventions) and Spec (stated requirements). |
Manual review, outside a ticket. Inside a ticket /implement applies the same rules inline — do not dispatch this from /ship.
The rubric lives in your context (## Review). This skill resolves what to review and handles the case where there is no spec.
| Argument | Change set |
|---|---|
| none | git diff main... — current branch against main. If the branch is main, git diff on the working tree. |
| a PR number or URL | gh pr diff <n> |
| a path or glob | git diff -- <path> |
--staged | git diff --cached |
git diff <resolved> --stat
cat prd/glossary.md 2>/dev/null || echo "NO_GLOSSARY"
Use glossary vocabulary in findings. Say "the Builder", not "the function in lib/default.nix".
gh pr view <n> --json body 2>/dev/null
The PR body links a ticket (Closes #n); read that ticket and its parent spec.
No PR, no ticket, no spec — run Standards only and say so in one line. Do not go hunting for a plausible spec, and do not infer one from the diff. A spec reconstructed from the code it is meant to judge always passes.
Conventions, naming, structure, imports, types, error handling, testing patterns, performance, security. Cite the repo rule you are applying — a convention visible elsewhere in the codebase, CLAUDE.md, or prd/decisions.md. A finding you cannot ground in a repo rule is a preference; drop it.
Tests get the same rule the code does. Flag any new test that asserts the toolchain works (install, build, compile, lint, file-exists, config-parses) or mirrors implementation. Flag a test with no behaviour in prd/behaviours.md and no public boundary behind it.
Wrong behaviour, unmet acceptance criteria, API mismatch, unintended side effects, scope creep. Scope creep counts: work the spec did not ask for is a finding even when the code is good.
Check the diff against the non-goals in prd/product.md too, if it exists. A change that quietly starts doing something the project decided not to do reads as a feature in review and as a commitment forever after.
Two lists, each ranked by cost of being wrong. State the bug, show the fix, stop. Skip nits.
Standards
1. <file:line> — <what is wrong> → <the fix>
Spec (or: skipped — no spec found)
1. <criterion> — <what is missing>
Fix what needs fixing when the user asks. Reviewing and fixing are separate calls.
Done when: both axes are run (or Spec is explicitly skipped with the reason), and every finding is either fixed or deferred with a reason. Checkable: re-run the diff and confirm no unfixed finding remains.