pr-feedback
Skill4 starsChanged 57 days agoMITView on GitHub
Answer the unresolved review feedback on a PR and clear the pr-feedback gate
15 other projects use a version of this.
Adopted by
Most recently taken up by
- RMATI-11861/Claude-Skillstheir own version, first seen Nov 16, 2026
- duncanmorris/agent-skillstheir own version, first seen Nov 13, 2026
- denormans/claude-code-workflowtheir own version, first seen Nov 8, 2026
- mmurakaru/skillstheir own version, first seen Nov 6, 2026
- krzysztofsurdy/code-virtuosotheir own version, first seen Nov 1, 2026
- mustafarslan/skillstheir own version, first seen Nov 1, 2026
Used with
Projects with it also use pr-create, pr-review and 2 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.
MinFEdevSkillpr-create
GitHub PR 생성 및 문서 자동 작성 스킬. 브랜치의 모든 커밋과 변경사항을 분석하여 체계적인 PR 문서를 생성하고 GitHub에 PR을 올림. 사용 시점: PR 올려줘, PR 생성해줘, PR 만들어줘 요청 시.
Also in 5 projects that use this
MicrosoftSkillpr-review
Verify the technical accuracy of a vscode-docs pull request against the VS Code source code in microsoft/vscode. Use when reviewing a docs PR for factual correctness: setting names, command IDs, default values, API shapes, keybindings, versioned availability, and described behavior.
Also in 5 projects that use this
SentrySkillcommit
Create commit messages following Sentry conventions. Use when committing code changes, writing commit messages, or formatting git history. Follows conventional commits with Sentry-specific issue references.
Also in 3 projects that use this
GoogleSkillcode-review
Adversarial systems, concurrency, memory-model, and architecture code review protocol. Use this skill whenever asked to review code, PRs, diffs, or branches in Filament.
Also in 3 projects that use this
Other versions15 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 15
- strands-agents/sdk-python★ 9.2ksame name
- strands-agents/harness-sdk★ 8.9ksame name
- jjanczur/tyran★ 79same name
- krzysztofsurdy/code-virtuoso★ 36same name
- BenjaminG/ai-skills★ 4same name
- hartleybrody/.agent-skills★ 4same name
- amknight/skills★ 2same name
- dversoza/claude-skills★ 2same name
- mustafarslan/skills★ 2same name
- denormans/claude-code-workflow★ 1same name
5 more not listed
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 repository10 other files it gives its agents.
- AGENTS.mdAGENTS.md
- agent-and-stateSkill
- babysit-prsSkill
- ci-and-releaseSkill
- development-automationSkill
- loggingSkill
- mode-blueprintsSkill
- project-mapSkill
- tui-standardsSkill
- web-frontendSkill
The file
We split this skill into the rules it gives an agent
3 rules. 0 are common, 1 is distinctive, and 2 are very specific to this project.
0
Common
Most of the 14 edited versions we read keep these.
1
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 120
If a finding is real but beyond this PR, say so in the reply and file it as an issue with the owning workstream: label rather than dropping …
- Very specific: line 123
Read advisory Codex findings as well as the enforced verdict.
- Distinctive: line 127
Never wait for a Codex completion marker.
“Common” is measured: kept by at least 6 in 10 of the 14 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.
Files
Work through the review feedback on a pull request until nothing is unanswered. Arguments (optional): the user’s arguments — a PR number; defaults to the current branch's PR.
Five things comment on a PR here and, until the pr-feedback gate existed, nothing
ever read any of it back. This is the procedure that reads it back. It is the one
place that job is written down; /ship and /babysit-prs
all point here rather than restating it.
Look —
make pr-feedback PR=<n>(oruv run python scripts/pr_feedback.py --pr <n> --jsonwhen you want to iterate). It prints every open item: unanswered findings per producer, unresolved human threads, and a requested-changes review.Three states are not failures and need no work:
pendingmeans the review has not run yet (wait for CI); a draft or a Dependabot PR is out of scope by design; and "Claude Review never posted a verdict" means the workflow did not run — checkgh run list --workflow "Claude Review"before touching the code, because nothing in the diff will fix it.- Read what was actually said —
gh pr view <n> --comments, andgh api repos/{owner}/{repo}/pulls/<n>/comments --paginatefor inline review comments. The verdict tells you how many findings are open; only the comments tell you what they are. Never answer a finding you have not read. - Decide fix or answer, per item. Fix when the finding is right. Answer when it is wrong, already handled elsewhere, or genuinely out of scope for this PR. Both are legitimate. Fixing everything to turn a check green is the worse failure: it lands unrequested changes on a branch someone already reviewed, and it teaches the reviewer that findings are cheap.
- Fixes — minimal change,
make test+make lint, commit with the repo's conventions and push to the same branch. Reply — one comment per producer, covering every finding, whichever way it went. A line each, saying what you did and why, ending with that producer's marker on its own line:
<!-- addressed: claude-review fixed=2 answered=1 -->fixed=counts the ones you changed;answered=counts the ones you argued. Two numbers rather than one because they are checked by different things: the reviewer's next pass reads the diff and re-reports a fix that is not there, and nothing but a person checks a disagreement. A bare<!-- addressed: claude-review -->is the older shape and still means "all of them, answered" — except from the PR's own author on the unattended lane, where "answered" is the one thing that account may not say, so write the counts.On an unattended PR a fix without this reply does not clear the gate. The next verdict reporting
open=0used to be the whole clearing mechanism, and it left the entire record of what a machine did about three findings as a number going down — which nobody can review.scripts/pr_feedback.pyreports that round asstopped being reported without anything saying what changed, and the way through is one comment. It is never capped and never expires.On a local branch none of it is required, for the same reason nothing else on that lane is: you did the fixing and you are the one merging. That covers the machine reviewer only — a human's unresolved thread and a
Request changesreview hold the check on every lane.The marker only counts if the comment is newer than the verdict it answers, so post it after the review, not before. If you push again afterwards, the next review pass will have read your reply and should stop reporting the finding at all.
A machine may not dismiss its own review. On an unattended PR, a
feature/issue-N-…branch, or a triage or sentinel branch,scripts/pr_feedback.pydiscards ananswered=claim from the PR's own author, so arguing a finding down there is inert and the gate stays red.fixed=from that same author is accepted, and the asymmetry is the point: a claimed fix is checked by the reviewer's next read, a claimed disagreement by nothing. Without it the applicant would be holding the key: the account that wrote the change also has write access. So on that lane the route through a finding you disagree with is to hand it back to a human — as a proposal, or viafeedback-override, which is a human's call and is recorded on the PR either way.AGENTS.mdprescribesfeature/<description>for human branches, so if you happen to name onefeature/issue-…you land in this rule too; rename the branch.Human threads — reply in the thread, then resolve it:
gh api graphql -f query='mutation($id:ID!){resolveReviewThread(input:{threadId:$id}){thread{isResolved}}}' -F id=<threadId>The thread ids are in the
--jsonoutput. Reply first, always — on an unattended PR that order is enforced rather than trusted: a thread the PR's own author resolved with no comment from them in it comes back as an open item. Resolved threads still accept comments, so the fix is a reply, not an un-resolve.- Confirm — re-run step 1 and report the final verdict with the PR URL. The
--jsonoutput carries aledger: one line per review round, what it found, and who wrote back to it. That is the shortest honest answer to "was this actually worked through, or did the number just go down".
For several PRs at once, or to run this unattended, use /babysit-prs fix — it spawns
the pr-responder subagent (.agents/reviewers/pr-responder.md) per PR in its own worktree.
Rules
- Never resolve a thread you did not answer, and never write an
<!-- addressed: -->marker without the reply that justifies it. The marker means "there is a reason below". Closing feedback silently is the behaviour this gate exists to stop — and on the unattended lane both halves are now checked in code rather than trusted. - Never claim a
fixed=you did not make. It is the one field the PR's own author may write on the enforced lane, and it is allowed only because the next review pass verifies it. Inflating it does not buy a merge; it buys a round. - A machine never answers its own review. On an unattended PR the only way to clear a finding is to fix it; disagreement is escalated to a human, never asserted. This one is enforced in code rather than trusted, because it is the single rule standing between an unattended merge and no review at all.
- The
feedback-overridelabel is a last resort and a human's call — for a gate that has genuinely gone wrong (a review that errored, a producer that changed format), not for feedback that is inconvenient. It is recorded in the sticky comment, so an overridden PR never looks like a clean one. - If a finding is real but beyond this PR, say so in the reply and file it as an issue
with the owning
workstream:label rather than dropping it.
Read advisory Codex findings as well as the enforced verdict. They are returned in advisory_items; a green gate does not mean those findings were addressed.
Native Codex findings use GitHub review threads, not producer verdict markers. Reply in the
original thread after fixing or explaining the finding; do not invent a Codex addressed marker.
Never wait for a Codex completion marker.
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.

