maintain-prove-it
Improve a repo's generated verification skill from what its past runs learned. Works on both kinds made by /create-prove-it, the permanent prove-<app> and a temporary prove-<migration>. Reads the receipts in .prove-it/runs/ (failed and blocked steps, retries, the commands that finally worked, steps that failed for reasons unrelated to the change, how each change was sorted into checks) and any bug that slipped past prove, then fixes drifted steps, writes in the working commands, adds the check that would have caught a slipped bug and a section for any new surface. For a temporary skill it checks whether the migration is done and, if so, suggests the end step. Never rewrites an answer key (spec, snapshot, old system) to match the app. Proves every edit live and reports product bugs without touching product code. Use for /maintain-prove-it, "update the verify skill", "the verifier is out of date", "prove missed this bug", "prove keeps failing on unrelated steps", or after a big merge or
ray-amjad · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add ray-amjad/skills --skill maintain-prove-it技能说明
name: maintain-prove-it description: Improve a repo's generated verification skill from what its past runs learned. Works on both kinds made by /create-prove-it, the permanent prove- and a temporary prove-. Reads the receipts in .prove-it/runs/ (failed and blocked steps, retries, the commands that finally worked, steps that failed for reasons unrelated to the change, how each change was sorted into checks) and any bug that slipped past prove, then fixes drifted steps, writes in the working commands, adds the check that would have caught a slipped bug and a section for any new surface. For a temporary skill it checks whether the migration is done and, if so, suggests the end step. Never rewrites an answer key (spec, snapshot, old system) to match the app. Proves every edit live and reports product bugs without touching product code. Use for /maintain-prove-it, "update the verify skill", "the verifier is out of date", "prove missed this bug", "prove keeps failing on unrelated steps", or after a big merge or migration milestone.
Maintain Prove It
A verification skill goes stale the day the app changes: a route moves, a button gets renamed, a new surface ships with nothing in the skill that reaches it. A stale skill is worse than none, because the next agent will trust it. Every prove run leaves a receipt that shows where the skill struggled. This pass reads those receipts and folds what they learned back into the skill.
Rules for the whole pass
- Only edit the skill's own folder (SKILL.md, scripts/). Never change product code here.
- When the app and the skill disagree, work out which one is wrong. The skill describing something the app no longer does is either drift (the app changed on purpose: fix the skill) or a product bug (the app broke: report it, don't rewrite the skill to match the bug). Read the commit that changed it. If you still can't tell, ask.
- Answer keys are not the skill's steps. The user's spec, the snapshots in
.prove-it/snapshots/and the old system a flow is compared with say what right looks like. A drifted step can be fixed, but an answer key that disagrees with the app is reported to the user, never rewritten to match. Only the user changes one. - Every edit you make to the skill gets proved live before you finish. An unproven edit is just more drift.
- Old verdicts are leads, not evidence. Earlier receipts, closed tickets and status docs tell you where to look. Only this pass's live runs count.
- Never loosen a check to make a flaky step pass. Record the failing run, then decide whether the app has a race or the check is ambiguous.
1. Find the skill
Look for prove-* folders under .claude/skills/ and .agents/skills/, at the repo root and in package folders. Prefer ones with the prove-it: marker comment. The marker says which kind it is: permanent skill (the app, one section per surface) or temporary migration skill (one migration, deleted at its end step). If there are several, ask which. If there are none, stop and suggest /create-prove-it.
2. Read the past runs
List the receipts in .prove-it/runs/, newest first. Start from the runs since the skill last changed (git log -1 --format=%cI -- <skill folder> gives the date). If the skill is uncommitted or there are only a few runs, read them all. For each receipt, note:
- Failed or blocked steps, and what the receipt says caused them.
- Retries and fixes: the command, port, selector or auth step that finally worked after the written one didn't.
- Steps that failed for reasons unrelated to the change. These are the clearest sign of drift: the change was elsewhere, but the skill's own step broke.
- Not covered lines that keep coming back, and the unlock each one names.
- Scope: the surfaces each diff reached. A changed path the Scope map doesn't list, or a surface the run drove with no section in the skill, means the map is behind.
- What the change touches: each flow the run listed as new or existing, its answer key, and any smoke test. A run that checked the whole app, skipped an existing flow behind shared code the change touched, or checked a new flow on parity alone when the change edited both sides, means the Sort checklist or the Scope map's shared-code rows need fixing.
- Answer keys used: new flows marked
no answer keyagain and again means specs aren't reaching the skill. Say so, and suggest where the user could put them (the PR body, a file in the diff).
If there are no receipts, say so. Then the only inputs are reported bugs (step 4) and a fresh run on the current change.
3. For a temporary skill, check whether the migration is done
Look at the evidence, not the plan. When every area in the Scope table is moved and the last run is green, suggest the end step now, before the PR that deletes the old system: the snapshots must be recorded while the old system still runs. Also suggest it if the old system's code is removed or about to be, or all traffic is on the new system. The end step is: save the cases as snapshots, add them to prove-<app>, prove them there, then delete the temporary skill. Don't run it without their OK. If the old system is already gone and the end step never ran, say so plainly: the cases can no longer be recorded from it, so the snapshots must come from the last receipts that ran it, or the new system as it is now, marked as the weaker key.
If it isn't done, update the Scope table's Moved column from the commits since the skill last changed.
4. Collect bugs that slipped past
If the user hasn't already said, ask in your first message, before the other work, whether anyone reported a bug that prove should have caught. For each one, find the receipt of the run that said PROVED on that change, and work out why it passed: the flow wasn't in scope, the step didn't look at the right thing, or the seed data hid it.
5. Sort what you found
Turn the notes into a
阅读完整来源 ↗