Respond Check

this skill

Screen review feedback and decide what deserves a fix. Every claim — from a colleague, from a bot, from a review command run earlier in this session — is tested against six gates and leaves with a verdict, its evidence, and a reply the reviewer can argue with.

This skill changes nothing. No edits, no commits, no comments posted, no threads resolved. Its entire output is a ledger and a recommendation; acting on it is the respond-action skill's job.

That is also why this skill can be routed to on the model's initiative while the respond-action skill and the respond-goal skill cannot: reaching for screening when someone pastes a review costs a report and nothing else.

Core thesis

A review is a set of claims, and a claim is not a work order. The expensive mistakes are symmetrical: a branch that fixes everything a reviewer says grows tests, guards, and comments that nobody can later justify; a branch that dismisses what a reviewer says ships the bug.

Screening separates the two before any code moves, and it separates them on the record — every verdict carries the evidence that produced it, so a reviewer who disagrees can attack the evidence instead of the judgment.

The screening question is never only "is this correct?" It is:

1. Is it true of the code as written?
2. Did this branch cause it?
3. Is it right for this project, against the decisions the project has already made and written down?
4. How often does the scenario it describes actually fire, and what happens when it does?
5. What would the fix cost everyone who reads the file afterward?

The rubric that answers these is references/screening-rubric.md and it is the authority; this skill collects the inputs and reports the outputs.

$ARGUMENTS contract

Non-flag text is the findings list. With no arguments and no --pr, look for review output already in this session; if there is none, ask for the feedback rather than inventing a review.

Phase 0: What this project has already decided

The alignment gate needs citations, so gather them before reading a single finding: AGENTS.md / CLAUDE.md and any nested equivalents covering the touched directories, .github/CONTRIBUTING.md, and the branch's own stated purpose — its pull request body, its commit messages, its ticket.

Resolve the provenance base here too (see the rubric's Provenance gate), and note the branch's push state with git status -sb.

A project with nothing written down has made no decisions to cite, which means the alignment gate will rarely fire. That is the correct outcome, not a reason to substitute a preference for a citation.

Phase 1: Collect

Gather from every channel references/feedback-sources.md describes that applies to this run: pasted text, this session's earlier review output, the pull request's three comment surfaces, and failing CI.

Normalize each claim into a finding record, split multi-claim comments into separate findings, drop the approvals and summaries that assert nothing, and merge duplicates — the same defect from four reviewers is one finding with four sources.

Carry forward the declines from any previous round in this session: a bot re-posting a finding that was already declined, against code that has not changed since, does not get screened twice.

Phase 2: Screen

Run the six gates on every finding, in the rubric's order, stopping at the first gate that settles it. Record each gate's evidence as you go — the diff that settled provenance, the citation that settled alignment, the trigger that settled odds. A verdict with no evidence behind it is not a verdict, and the report must not contain one.

Two disciplines while screening:

• Read the code, do not trust the claim. Automated reviewers assert with total confidence regardless of whether they are right, and three bots agreeing is one opinion repeated, not corroboration.
• Cost the smallest fix that works, not the one the reviewer proposed. Reviewers routinely propose a mechanism where a condition would do, and the cost gate should judge the cheaper option.

Use ask-user-choice for ask verdicts only — findings where the truth depends on intent only the author has, or where the alignment citation is genuinely contested. Everything else is decided here.

Phase 3: Report — the output contract

1. Hero block (1–3 lines): N to fix, M deferred, K declined plus the branch name and the feedback sources screened.
2. ## Ledger — one entry per finding: id, source (author, and whether it is a bot), the claim quoted, location, the gate results with their evidence, and the verdict with its reason. Merged duplicates list all their sources on one entry.
3. ## To fix — the accepted findings in the order they should land, each with the minimal fix and whether it wants a forward commit or a fixup!.
4. ## Deferred & declined — with the follow-up recommendation or the drafted reply. A declined improbable reply states the trigger the scenario needs, so a reviewer who knows a caller that produces it can say so.
5. ## Ledger key — branch, HEAD SHA, and the feedback digest. This is what the respond-action skill checks to know the screening still describes the current tree.
6. End with an ask-user-choice panel: run the respond-action skill on the accepted findings, run it with --reply so the drafted replies get posted too, re-screen with the deferred findings opted in, or stop. In a non-interactive run (CI, subagent) record the panel's options in the report and stop.

Nothing in this phase edits, commits, or posts. If the run reaches a point where a fix seems obvious enough to just do, that is the failure mode this skill exists to prevent — write it in ## To fix and hand off.

Portability notes

• ask-user-choice — present the listed options and wait for the user to pick one. Hosts with a structured multiple-choice tool (Claude Code's AskUserQuestion) should use it; otherwise print a numbered list and wait for a numbered reply. Never proceed on an assumed answer.
• $ARGUMENTS — the text the user passed when invoking this skill. If your host does not substitute it, read it as the user's request in the current turn, and ask when there is none.
• Bundled files — every relative path in this skill points at a file shipped inside this skill directory. Read them from here, not from the host's plugin tree.