agentleFS
stijnvanhullestijnvanhulle/agents/.agents/skills/review/SKILL.md

review

Skill2 starsChanged 56 days agoMITView on GitHub

Reviews or ships a GitHub pull request or GitLab merge request. Review mode reads the request, runs the conventions, deslop, jsdoc, humanizer, and documentation passes, then asks what to do with each finding and with the request: comment, request changes, approve, fix, or close. Ship mode takes your own branch to merge-ready, including branch name, checks, changeset, commits, title, body, push, and CI. Use when asked to review a PR or MR or whether one is ready to merge, and load it before the first commit whenever work will end in a commit, push, or PR, including a session told to push to a branch.

  • Commits and pushes

192 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use close-spec, design-with-images and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • dzhngSkill

    close-spec

    When a spec under specs/<feature>/ is done shipping (a write-spec build, or any planned task that produced a spec), archive it to specs/done/ and rewrite it from a build-plan into a durable rationale — the why, the principles, the invariants — pointing back to the real code for the how. Use when implementation has landed and the plan no longer matches what shipped, or the user says a feature/spec is finished. Pairs with [write-spec](../write-spec/SKILL.md) (the plan this closes) and [review](../review/SKILL.md) (run before closing).

    Also in 5 projects that use this

  • dzhngSkill

    design-with-images

    Explore UI and visual design with image generation, then iterate the real implementation against the user-selected concept. Use when brainstorming visual directions, offering mockup variants, or implementing an approved generated design; do not stop at the mockups or the first code approximation.

    Also in 5 projects that use this

  • LambenthanSkill

    daily-arxiv

    每日从 arXiv 拉取新论文,过滤相关性,auto-ingest 高优先级论文,检测 SOTA 更新

    Also in 5 projects that use this

  • LambenthanSkill

    exp-eval

    实验判决门:Review LLM 独立评判实验结果 → 4 种判决路径 → 自动更新 claims confidence、ideas status、graph edges

    Also in 5 projects that use this

  • +3 more from dzhng
  • +5 more from Lambenthan

The file

Show its rules and text

We split this skill into the rules it gives an agent

14 rules. 0 are common, 7 are distinctive, and 7 are very specific to this project.

  • 0

    Common

    Most of the 156 edited versions we read keep these.

  • 7

    Distinctive

    This author's own choices. Worth a look.

  • 7

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read the title, body, linked issue, commits, open review threads, CI status, and the diff.

    line 21
  2. Distinctive:

    Write the diff to a file outside the working tree for the subagent.

    line 24
  3. Distinctive:

    Check the claims in the description and commits against the diff.

    line 26
  4. Very specific:

    Write the findings list (see Output), then in the same reply follow the ask skill:

    line 48
  5. Very specific:

    After those answers, one question for the verdict, since it depends on what stays: Comment, Request changes, Approve, or Close.

    line 53
  6. Very specific:

    Run humanizer over anything you post, since you wrote it in this task.

    line 69
  7. Very specific:

    Never commit to main. If git branch --show-current is main, branch from an up-to-date main with the branch skill's …

    line 75
  8. Very specific:

    When the session or harness assigns the branch, such as a cloud session told to push to claude/<name>, keep that name: the assignment wins, …

    line 79
  9. Very specific:

    Run the changeset skill for a change to a versioned package.

    line 98
  10. Very specific:

    If the work is already committed and a commit mixes changes, split it before the first push: git reset to the merge base with origin/main, …

    line 107
  11. Distinctive:

    After a push, add new commits instead.

    line 109
  12. Distinctive:

    Before asking to push, confirm that:

    line 131
  13. Distinctive:

    Run any step that did not run.

    line 140
  14. Distinctive:

    After shipping: the PR URL, then one bullet per gate item that was skipped, or "nothing skipped".

    line 171

“Common” is measured: kept by at least 6 in 10 of the 156 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Review

Two modes for a pull request or merge request. Review judges a request someone opened, or one you want checked before merging, and acts only on what the user picks. Ship takes your own branch to merge-ready, in order. Load this skill before the first commit, since the branch name and the commit split are hard to change after a push.

Review

1. Fetch the change

Pick the host skill, github or gitlab, from the input or the remote, and use its commands for every read and write in this mode. The input is #12, !12, owner/repo#12, a request URL, or nothing for the current branch's request.

Read the title, body, linked issue, commits, open review threads, CI status, and the diff. When the working tree is clean, check out the branch so the review sees the connected code and can run the checks. Otherwise fetch the branch and read files with git show origin/<branch>:<path>. Write the diff to a file outside the working tree for the subagent.

Check the claims in the description and commits against the diff.

2. Run the passes

In Claude Code and Cursor, run the code-reviewer subagent with the diff file and the checked-out branch, and take its findings for passes 1 to 5. Elsewhere, run the passes yourself, substance before style before prose, reading each skill for its criteria:

  1. conventions and correctness: logic errors, edge cases, security at trust boundaries, and a test for every code change.
  2. deslop: the reuse-first ladder and the code tells.
  3. jsdoc: every touched export.
  4. humanizer: changed comments and markdown, plus the title and body.
  5. documentation: a changed docs page or blog post.
  6. Ship steps 1 to 7 as checks, so the branch name, changeset, commit split, body, and CI get a pass or fail.

Skip a pass with nothing in its category rather than forcing a finding. Until step 4, nothing is edited, posted, or pushed.

3. Report, then ask

Write the findings list (see Output), then in the same reply follow the ask skill:

  • One question per finding, named after it (Unused retry flag in client.ts?), offering Comment, Fix, or Skip. Comment posts it on the request, Fix edits the checked-out branch. Send a follow-up call when there are more than four.
  • After those answers, one question for the verdict, since it depends on what stays: Comment, Request changes, Approve, or Close. Recommend Request changes with a blocking finding, Approve with none, and Comment otherwise. GitHub refuses Approve and Request changes on your own request, so offer Comment, No review, or Close there.

A skipped question stops that part: never post, approve, or close on a guess.

4. Act on the answers

ChoiceAction, with the host skill's commands
Comment, Request changes, ApproveOne review with the body and each confirmed finding as an inline comment. On GitLab the note carries a request for changes
ClosePost the reason, then close
FixFollow deslop step 4 on the checked-out branch: surgical edits, format, lint, tests, re-audit. A merge conflict gets its own merge commit before the fixes. The fixes stay in the working tree, and a commit or push is a separate ask as Ship step 6 says

A posted finding carries its path:line, the problem, and the fix, in one or two sentences. Run humanizer over anything you post, since you wrote it in this task.

Ship

1. Branch

Never commit to main. If git branch --show-current is main, branch from an up-to-date main with the branch skill's <category>/<ISSUE-REF>_<slug> shape, leaving out the reference when there is none.

When the session or harness assigns the branch, such as a cloud session told to push to claude/<name>, keep that name: the assignment wins, and a rename breaks its push. Otherwise a branch that does not match the shape, such as a desktop worktree's claude/… branch, gets renamed with git branch -m before its first push. Keep the name of a branch already pushed.

2. Checks

From the repo root, run every format, lint, typecheck, and test script package.json defines, and the build after package-source changes:

pnpm format && pnpm lint:fix && pnpm typecheck && pnpm test

A single package's tests or a lint over changed files do not replace the full run. Skip a script the repo does not define. Fix failures at the source.

3. Changeset

Run the changeset skill for a change to a versioned package. Docs, CI, and tests alone need none.

4. Commits

One imperative Conventional Commit per logical change, under 72 characters, no period. Review git diff --cached first. Never commit secrets or generated output, and generate lockfiles with pnpm.

If the work is already committed and a commit mixes changes, split it before the first push: git reset to the merge base with origin/main, then stage and commit each change on its own. After a push, add new commits instead.

5. Title and body

The title is an imperative Conventional Commit under 72 characters with no period. Take the type from the change or the issue label, not feature or hotfix. Put Closes #123 in the body, and open the issue with the issue skill when none exists.

Fill .github/pull_request_template.md when it exists, without deleting sections. Otherwise use:

  • Changes: what changed and why, in 1 to 3 sentences.
  • How to test: three steps from a clean checkout, ending in the expected result. When the diff cannot supply them, use the ask skill. Add screenshots for visible changes.
  • Checklist: tick only what ran, and explain each unticked item.
  • Release impact: the changeset, docs-only, and breaking boxes match the diff.
  • Impact: who is affected, and any migration.

Match the format of recently merged PRs in the repo. Keep each section short: when the body runs long, cut background, not sections. Run humanizer over the prose.

6. Gate, then push

Before asking to push, confirm that:

  • the branch name has the shape, or the branch was assigned or already pushed
  • the full check scripts ran and passed
  • the changeset skill ran, or the diff needs no changeset
  • each commit holds one logical change
  • deslop ran on the code and humanizer on the PR body
  • the diff stays in scope: no drive-by refactors and no force-push

Run any step that did not run. When you cannot, stop and name the skipped steps instead of pushing.

Pushing publishes the commits, so follow the ask skill first: name the branch and offer push and open the PR, push only, or stop. A yes covers that branch only, and main is never pushed. On stop, leave the commits local and say so.

git fetch origin main && git pull --ff-only
git push -u origin <branch>

Then open the request with the host skill, github or gitlab, ready for review and with existing labels only. One PR does one thing.

7. CI and review

Fix red CI before handoff, and re-run only infrastructure failures. A blocked merge state with green CI means a required review is missing. Say so, since a review is not yours to supply. Reply to each review comment with what changed and how to verify it, and propose before making a large requested change.

Output

For a review: one bullet per finding, grouped by severity (blocking, should-fix, nit) within pass order: path:line, the pass, the fix, and Comment, Fix, or Skip as your recommendation. Then one row per Ship step with pass or fail and the reason. A miss that cannot change after the push is a fail marked "left as is". End with one bullet naming what you checked and found fine, including CI and mergeability, and the verdict: ready, or what blocks it. After acting: the review or comment URL, or "not posted", and one bullet per fix left in the working tree.

After shipping: the PR URL, then one bullet per gate item that was skipped, or "nothing skipped".

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.