build-review-loop
Skill0 starsChanged 50 days agoMITView on GitHub
Build the user's own issues in parallel with one builder agent and worktree per issue, then self-review each draft pull request, post the findings, send them back to the same builder, and re-review until every current head has no actionable findings. Use when asked to build issues with a self-review and fix loop before the user's final review. Single tasks use code-build; review-only work uses code-review.
2 other projects use a version of this.
Adopted by
Most recently taken up by
- UnfoldCRO-Organization/cro-agent-skillstheir own version, first seen Nov 16, 2026
- kl-autonomous/agent-team-skillstheir own version, first seen Oct 20, 2026
Other versions2 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 2
- kl-autonomous/agent-team-skills★ 2same name
- UnfoldCRO-Organization/cro-agent-skills★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository49 other files it gives its agents.
- AGENTS.mdAGENTS.md
- accessibility-auditSkill
- ai-tools-updateSkill
- antickSkill
- app-pentestSkill
- branch-statusSkill
- ci-cd-triageSkill
- code-auditSkill
- code-buildSkill
- code-debugSkill
- code-hardeningSkill
- code-refactorSkill
- code-reviewSkill
- code-securitySkill
- commit-messageSkill
- conciseSkill
- create-issueSkill
- database-auditSkill
- deep-wiki-htmlSkill
- dependency-auditSkill
- dependency-override-cleanupSkill
- dependency-upgrade-majorSkill
- dependency-upgrade-minorSkill
- documentation-generationSkill
- electron-appSkill
- github-actions-lockSkill
- github-actions-security-auditSkill
- interviewSkill
- make-senseSkill
- new-skillSkill
- performance-auditSkill
- plan-reviewSkill
- pr-descriptionSkill
- private-desktop-releaseSkill
- public-desktop-releaseSkill
- reclaim-dev-spaceSkill
- save-omarchy-window-layoutSkill
- setup-vercel-analyticsSkill
- simplifySkill
- skill-reviewSkill
- soc2-readinessSkill
- spec-statusSkill
- submit-reviewSkill
- sync-branchSkill
- system-storage-auditSkill
- test-suite-hardeningSkill
- update-skillsSkill
- use-chromeSkill
- vercel-domain-cleanupSkill
The file
We split this skill into the rules it gives an agent
10 rules. About 0 are common, 7 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈7
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Distinctive: line 24
Read each issue and its discussion.
- Very specific: line 30
Keep a compact ledger with one entry per issue: builder handle, worktree, branch, pull request link, latest head SHA, review round, open …
- Distinctive: line 37
Use the host's delegation and worktree tools to give each issue its own builder, worktree, and branch, based on the latest fetched base …
- Distinctive: line 59
When a builder returns an open decision, ask the user, then send the answer back to the same builder.
- Distinctive: line 62
Start independent builders together and review finished pull requests while others are still building.
- Very specific: line 67
Run code-review completely on each pull request: the whole base-to-head change against its exact issue, not the builder's summary.
- Very specific: line 74
After code-review, run any user-Chrome checks the builder returned as unrun yourself with use-chrome, one at a time, within the user's …
- Distinctive: line 91
If a builder disputes a finding, check the evidence against the requirement and settle it on technical grounds.
- Distinctive: line 98
When a builder reports a fix, verify the pushed head and repeat step 3 on the whole pull request: recheck every earlier finding and look …
- Distinctive: line 110
Before announcing completion, confirm every issue has a pull request whose latest head matches a clean verdict.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Build Review Loop
Take each requested issue through building, a draft pull request, review, published findings, fixes, and re-review. The coordinator (you) owns every review verdict; each issue's builder owns its implementation and branch. These are the user's own tasks, so every review is a self-review.
Authorization
Invoking this loop authorizes builder delegation, commits, pushes, draft pull
requests, and publishing findings with submit-review in self-review mode.
The user keeps the final review: do not merge, deploy, approve, mark a pull
request ready, or change ticket or project status. The repository's other
approval rules still apply. Continue independent work while a decision is
pending.
1. Map issues to builders
Read each issue and its discussion. Capture reproduction steps, acceptance criteria, dependencies, and scope boundaries. Ask only about missing information that blocks correct work. Spot shared files, colliding migration numbers, and ordering dependencies early, and coordinate them without merging the issues into one implementation.
Keep a compact ledger with one entry per issue: builder handle, worktree,
branch, pull request link, latest head SHA, review round, open findings,
verification results, and state (queued, building, reviewing, fixing,
clean, or blocked). A blocker is never clean.
2. Start one builder per issue
Use the host's delegation and worktree tools to give each issue its own builder, worktree, and branch, based on the latest fetched base branch unless the user names another. Verify each worktree's base before work starts. Queue issues beyond the available concurrency, keeping one builder per issue. Keep any provider or model settings the user asked for, and say when one is unavailable.
Give each builder a self-contained task with:
- the exact issue URL, scope, dependencies, and shared-file notes;
- an instruction to use
code-build, including its meaningful commits, behavioral tests, focused checks, and draft pull request linked to the issue; - the user's earlier consent or refusal for expensive checks;
- an instruction to return the pull request URL, pushed head SHA, changed behavior, checks run with results, unrun checks, and open decisions;
- an instruction to implement only and leave review to the coordinator;
- an instruction to return any question for the user, from
code-buildorsync-branch, such as a layout sketch for a new screen or a conflict where both sides cannot be kept, as an open decision instead of waiting for the user; - an instruction not to use the user's main Chrome, and to return any browser check that needs it as unrun, with the steps to perform.
When a builder returns an open decision, ask the user, then send the answer back to the same builder.
Start independent builders together and review finished pull requests while others are still building. Keep each builder's handle for later fixes.
3. Review every finished pull request
Run code-review completely on each pull request: the whole base-to-head
change against its exact issue, not the builder's summary. Record the reviewed
head SHA and findings. A builder's own "done" or "clean" claim, green CI, and
passing tests do not establish a clean review. If the head changes during the
review, review the new head before recording a verdict; a verdict applies only
to the SHA actually reviewed.
After code-review, run any user-Chrome checks the builder returned as unrun
yourself with use-chrome, one at a time, within the user's consent for
browser checks. Record the results, and treat a failed check as a finding for
that pull request.
4. Post findings and return them to the builder
For each pull request with actionable findings, run submit-review in
self-review mode: inline comments without a username tag, a comment review, and
the pull request kept as draft. Record the review link.
Send the exact findings and link to that issue's existing builder, resuming it
in the same worktree and branch to fix, test, commit, and push to the same pull
request. For a merge-conflict or behind-base finding, tell the builder to bring
the branch up to date with sync-branch, merging the pull request's base
branch, then re-review. Keep other builders on their own issues.
If a builder disputes a finding, check the evidence against the requirement and settle it on technical grounds. Ask the user when intended behavior is ambiguous. Never drop a valid finding to end the loop. Persistent permission, environment, or capability failures are blockers to report.
5. Re-review after every fix
When a builder reports a fix, verify the pushed head and repeat step 3 on the whole pull request: recheck every earlier finding and look for regressions the fix introduced. Post new or remaining findings through step 4.
Repeat until the current head has zero actionable findings. Do not stop after a fixed number of rounds. If an issue makes no progress across rounds, investigate or escalate that blocker while other issues continue. Any push, sync, or conflict resolution after a clean verdict invalidates it and needs a new review.
6. Hand back the batch
Before announcing completion, confirm every issue has a pull request whose latest head matches a clean verdict. Report overlapping files, migration collisions, and merge-order constraints across the batch; do not merge one pull request to fix another.
Return one row per issue:
| Issue | Pull request | Reviewed head | Rounds | Final state |
Then summarize fixes made, passing checks, unrun checks, open decisions, and integration conflicts. Say clearly whether all pull requests have no actionable findings and await the user's final review, or which are blocked. Send short progress updates while work continues; a progress update is not completion. CI monitoring and repair are extra work only when requested.
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
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.

