agentsclimarketplace

Miki review loop

Skill mikiwiik/claude-code-skills/miki-review-loop

Personal Claude Code skills (miki-*), symlinked into ~/.claude/skills/ via install.sh.

Install
npx -y skills add mikiwiik/claude-code-skills --skill miki-review-loop

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Miki's personal pre-self-review loop. Iteratively runs /review on the current PR and addresses obvious issues until only minor nits remain or 3 iterations have run. Surfaces judgment calls back to the user, makes atomic commits per project convention, pushes once at the end. Invoke when the user says "/miki-review-loop", "loop the review", "iterate on the review", or similar after a PR has been created.

SKILL.md

7.7 KB, ~1.8k tokens by cl100k_base, as published. Nobody here has run it

miki-review-loop

What this does

Runs /review → addresses issues → re-runs, up to 3 iterations. Exits when only nits remain, a judgment call surfaces, or the cap is hit. Ends with a written summary of every change and commit.

Preconditions

Verify in order; stop and report on first failure:

  1. PR identification.
    • Argument given (bare 84, #84, PR 84, PR #84, or full PR URL): normalize to a bare number, run gh pr view <n> --json number,title,headRefName,baseRefName,state, announce Processing PR #N — "<title>" on <head> → <base>, proceed.
    • No argument: derive from current branch via gh pr view --json number,title,headRefName,baseRefName,state. Announce Detected PR #N — ... and ask to confirm. If ambiguous, list candidates and ask.
    • No PR or not OPEN: stop.
  2. Clean working tree (git status --porcelain empty).
  3. Not on main / master.
  4. Verification command available (see below). If not, follow the prompt-to-create flow before iteration 1.

Verification command

The loop runs the repo's verification command between iterations to confirm fixes don't break the build. Detect on each run:

  1. package.json scripts — try in order: check, verify, validate, ci, test. For each, prefer the :ci variant when present (e.g. check:ci over check) — tuned for non-interactive runs. First match wins.
  2. Other stacks — Cargo.toml → cargo check && cargo test; Makefile with check target → make check; pyproject.toml with [tool.<runner>] → obvious check/test task.
  3. CLAUDE.md — grep for a documented verification command (e.g. under a "Verification" header or a fenced block tagged as the canonical check). If present, use it.

One candidate found → use it, tell the user. Multiple plausible → list and ask.

If none can be found

Stop before iteration 1 and tell the user:

I couldn't find a verification command for this repo. The loop runs one between iterations to catch broken intermediate commits before they compound.

Options: (a) Add a check script to package.json aggregating lint + typecheck + test. (b) Document the command in CLAUDE.md under a "Verification" section. (c) Give me the command now — for this run only, won't persist. (d) Skip verification — not recommended; risky on iteration 2+ (broken intermediate commits compound).

For (a) or (b), draft the addition and show it to the user before writing; persist as a separate atomic commit via the user's normal flow.

Auto-merge handling

After preconditions pass, check gh pr view <n> --json autoMergeRequest,reviewDecision. If auto-merge is enabled, disable it (gh pr merge <n> --disable-auto) before iteration 1 and record the merge method (SQUASH/MERGE/REBASE) — otherwise the merge can fire on the pre-fix SHA the moment CI greens. Announce auto-merge paused for loop duration; also mention if reviewDecision is APPROVED (that's why the pause matters).

Restore via gh pr merge <n> --auto --<method> after the final push, and on every exit path: clean (only nits or no findings at all, no commits made), judgment call surfaced, iteration cap reached, build verification failed, push failed. The trivial-clean exit where no commits were made still requires restoration — don't conflate "nothing to push" with "nothing to restore". Restoration is the last action before writing the summary, not something the user has to trigger by saying "please merge". If disable or restore fails, stop and report — don't iterate without the pause, don't leave auto-merge in a different state than you found.

Severity bands

Classify each /review finding:

  • Blocker — breaks correctness, security, or build. Fix.
  • Major — clear defect or convention violation, single right answer (e.g. missing null check on documented-non-null field). Fix.
  • Minor — single obvious right answer, no behavior change (e.g. unused import, typo). Fix.
  • Nit — style/preference, multiple valid choices. Skip.
  • Judgment call — depends on intent, tradeoffs, or info not in the diff (architecture, naming, public-API contracts, behavior changes, anything needing a new test). Stop and ask.

When in doubt between Minor and Judgment call, treat as judgment.

The loop

For iteration i in 1..3:

  1. Run /review.
  2. Classify findings; show the grouped list to the user before acting.
  3. Any judgment calls → stop, ask one by one. Don't silently skip.
  4. No Blocker/Major/Minor remain → exit successfully (only nits, or nothing). Restore auto-merge before writing the summary (see Auto-merge handling).
  5. Address Blocker + Major + Minor. If a finding turns out more ambiguous while fixing it, stop and ask.
  6. Run the verification command. If it fails, fix the failure and re-verify before committing. If you can't fix cleanly within the same logical change, surface as a judgment call.
  7. Commit atomically (see below).
  8. Track every change (file path + one-line) and commit (sha + message) in a running log.

If iteration 3 finishes with non-nit issues remaining, stop and report.

Commit and push cadence

  • Atomic commits per project convention (Conventional Commits in most projects — check git log or CLAUDE.md). Two unrelated findings = two commits.
  • Amend only when the new edit is part of the same atomic change as the previous commit — verification fix-the-fix, a follow-up iteration completing an incomplete fix. Never across two separate findings.
  • Never amend a commit already pushed (prior loop run or anything not made by this loop). Treat existing history as immutable.
  • Push once, at the end of the loop. Single git push, no matter how the loop terminates. Never force-push.

Final message format

## Review loop summary

**PR**: #<number> — <title>
**Iterations run**: <n>/3
**Exit reason**: <clean (only nits remain) | iteration cap reached | judgment call surfaced | precondition failed | build verification failed | no verification command>

### Changes made
- <file:line> — <one-line description>
- ...

### Commits in this review cycle
- <sha> — <commit message>          (new | amended)
- ...

### Remaining issues
- **Nits** (skipped): <count + one-line list>
- **Judgment calls** (need input): <list with the question, or "none">
- **Unaddressed non-nits** (only if cap hit): <list>

### Pushed
- <yes / no — if no, why>
- **Auto-merge**: <paused → restored as <method> | unchanged | restore failed: <reason>>

List every commit produced, including ones later amended (mark as amended). If no changes were made, say so plainly rather than printing empty lists.

Operating rules

  • Never resolve findings by deleting tests, weakening assertions, or adding eslint-disable / @ts-ignore. Those are judgment calls.
  • Never broaden scope. If /review flags untouched code, surface as a judgment call ("fix here, separate PR, or skip?") rather than silently expanding the diff.
  • Stay in the diff. Clean up what this PR introduced, not the surrounding area.
  • If /review fails or returns nothing parseable, stop and report — don't guess what it would have said.

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.

Gives 0 of the 12 instructions most review quality skills give in ~1.8k tokens

Counted across 1,048 of the 1,783 authors here whose files we hold, read 2026-08-07

  • ask questions one at a timein 81 of 1048, across 64 files
  • provide a recommended answer for each questionin 73 of 1048, across 50 files
  • explore the codebase instead of asking answerable questionsin 66 of 1048, across 42 files
  • resolve dependencies between decisions one-by-onein 42 of 1048, across 17 files
  • interview the user relentlessly about the planin 38 of 1048, across 13 files
  • order findings by severityin 31 of 1048
  • resolve each branch of the decision treein 27 of 1048, across 5 files
  • run a grilling sessionin 26 of 1048, across 5 files
  • update CONTEXT.md immediately when a term is resolvedin 26 of 1048, across 11 files
  • propose precise canonical terms for vague languagein 25 of 1048, across 7 files
  • create documentation files lazilyin 24 of 1048, across 5 files
  • Assign severity to every findingin 24 of 1048

Said here and by no other author read

  • verify preconditions before starting the loop
  • run review up to three times
  • skip all nit level issues
  • run verification command between iterations
  • push once at the end of the loop
  • restore auto merge settings before writing summary

Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.