review-revision-candidates
Skill3 starsChanged 54 days agoMITView on GitHub
Run a revision round on a candidate a reviewer rejected: re-read the governing files at the start of the round, treat the numbered list as the entire change set, and deliver bytes plus anchors rather than prose about the rejection. Use when a reviewer returns a patch, draft or page edit as REJECTED with items to fix. Not for open-ended rewrites or for reviewing someone else's work.
In this repository59 other files it gives its agents.
- AGENTS.mdAGENTS.md
- agent-harness-settings-paritySkill
- agent-operating-contract-evolutionSkill
- ai-coding-agent-delegationSkill
- arch-improvement-planSkill
- contract-bound-deliverablesSkill
- cross-agent-repair-routingSkill
- define-goalSkill
- delegated-work-acceptanceSkill
- design-an-interfaceSkill
- design-mdSkill
- discord-rich-messageSkill
- dotnet-repo-workflowSkill
- evidence-backed-analytical-writingSkill
- evidence-boundary-disciplineSkill
- find-nearbySkill
- gddSkill
- generated-runtime-repairsSkill
- ghc-searchSkill
- git-auto-sync-pipelinesSkill
- github-code-reviewSkill
- github-repo-managementSkill
- grounded-citationsSkill
- harness-engineeringSkill
- hermes-context-file-loadingSkill
- hermes-modelsSkill
- hermes-updatesSkill
- hoa-introspect-distillSkill
- hoa-introspectSkill
- hoa-skill-reviewSkill
- hoa-skill-workSkill
- html-artifactSkill
- jupyter-live-kernelSkill
- kanban-workflowsSkill
- m365-copilot-cdpSkill
- measurement-campaign-evidenceSkill
- microsoft-officeSkill
- notes-to-prdSkill
- ocr-and-documentsSkill
- official-document-verificationSkill
- omp-headless-orchestrationSkill
- ops-claim-verificationSkill
- package-release-verificationSkill
- plan-to-issuesSkill
- project-phase-closeoutSkill
- report-claim-auditSkill
- research-stream-consolidationSkill
- rss-feedsSkill
- skill-library-curationSkill
- skill-porting-and-publishingSkill
- spec-convergenceSkill
- structural-validity-gatesSkill
- subagent-driven-developmentSkill
- systematic-debuggingSkill
- teams-cdpSkill
- tmux-agent-controlSkill
- whisperSkill
- writing-plansSkill
- writingSkill
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 4 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
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 20
Use when a reviewer returns a candidate (patch, draft, page edit) as REJECTED with a numbered list and asks for vN+1, and the tree must …
- Distinctive: line 39
Every numbered item is a gate.
- Very specific: line 41
Delete hunks the review excluded even when they look harmless — metadata fields (updated:, last_accessed:) are commonly excluded in …
- Distinctive: line 73
Absence findings need the broader search before they ship: Search the abbreviation and the full phrase, the section number and the subject …
- Distinctive: line 77
Attribute access failures to the run, not the resource: Write "this client, this date, this path returned 403", never "the source is …
- Distinctive: line 98
Do not restate the review back at the reviewer, and do not send an acknowledgement between rounds — the next message is the candidate.
≈ “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
Revision rounds on a rejected candidate
Use when a reviewer returns a candidate (patch, draft, page edit) as REJECTED with a numbered list and
asks for vN+1, and the tree must stay untouched while that happens. The deliverable of a revision round
is bytes plus anchors, never prose about the rejection.
1. Re-read the governing files at the start of every round
- The review/request file is amended in place between rounds: status line, verified-evidence tables, and the revision advice all change. A plan built on the earlier read is stale, and arguing against a withdrawn instruction is the fastest way to lose the round. Read the review file, the frozen request, and that directory's evidence/ledger files again before editing.
- Confirm the baseline has not moved:
git rev-parse --short HEAD, then check the targets are still clean (git diff --stat -- <paths>). Peers publish during the round; a patch verified earlier may no longer apply. - Note whose working-tree changes are present before you report anything about the tree. A dirty
git statusunder the reviewer's own records directory is theirs, not yours; say so explicitly rather than letting your "repo untouched" claim look false.
2. Treat the numbered list as the sole change set
- Every numbered item is a gate; nothing adjacent gets fixed on your own initiative, however obvious. Out-of-scope improvements cost a round and read as not listening.
- Delete hunks the review excluded even when they look harmless — metadata fields (
updated:,last_accessed:) are commonly excluded in candidate rounds. Do not argue for them; report that they were dropped and that the rest does not depend on them. - Convert each item into something you can grep for afterwards: the banned phrase, the required sentence, the mandatory locator.
3. Versioning: the rejected candidate is immutable
- Build
vN+1from pristine copies of the target files, in scratch, and write new files (vN+1-<name>.patch, plus the edited whole pages). Never overwrite or "fix" the rejected patch: the reviewer may hold its bytes, and an in-place edit destroys the ability to compare versions. - Apply edits as exact unique-anchor replacements, asserting each anchor occurs exactly once and writing only after every replacement matched (all-or-nothing). A partial application produces a candidate that looks right in the diff you read and is corrupt where you did not look.
- Produce diffs with stable labels so they apply at the repo root:
diff -u --label a/<path> --label b/<path> <pristine> <edited> > vN-<name>.patch. - Per version, report numstat and
git apply --checkrun from the repo root against current HEAD. A clean apply is a structural pass only — say so; never present it as content acceptance.
4. Pre-resubmit self-review
- Grep the new candidate for every phrase the review banned and for its softened variants. Reviewers reject a conclusion, not a sentence: a rewrite that trades "is closed" for "is effectively closed" or "no longer" reintroduces the same over-claim. Check the inverse direction too — a banned "never X" is usually re-smuggled as "only if Y".
- Downgrade cross-caliber inferences to the observation the evidence supports. A later, differently defined measurement does not retire an earlier one: state the new dated value with its measure, keep the older figure with its own date and source, and say the definitions are not aligned instead of ranking or converting them.
- Do not relabel inherited numbers to make them comparable. Preserving an earlier team's figure "as headcount" is itself an unverified definition claim. Mark it prior/attributed, name what is unverified, and refuse the conversion.
- Absence findings need the broader search before they ship. Search the abbreviation and the full phrase, the section number and the subject matter; then disclose a found counterexample and scope the limit to the inspected versions and dates. "Not in the version I read" is the claim; "not in the law" is not yours to make.
- Attribute access failures to the run, not the resource. Write "this client, this date, this path returned 403", never "the source is unavailable" or "the channel is closed"; and keep the four things a visibility question really contains separate (index/link state, access gate, alternative public copies, whether the document was ever filed or required).
- Re-check the mechanical leftovers: no excluded metadata hunk, no forbidden phrase, valid frontmatter
(
yaml.safe_loadon the block), and a diff line count that matches what you report.
5. Deliver: bytes plus an anchor table, not a report
One reply carrying:
- Paths and per-file numstat + apply-check result, plus the baseline SHA they were built against.
- A short mapping "review item N → what changed", quoting the new sentence fragment, so the reviewer can verify without opening the patch.
- An evidence table, one row per load-bearing claim: claim → source URL and exact version → figure/section/line locator → short verbatim quote → caliber and status limit (proposed vs enacted, authorized vs actual, aggregated vs per-unit).
- What stayed unresolved, named as unresolved, with the reason the search could not settle it.
- The scope statement: not applied, no write lease taken, no metadata or adjacent files touched, and what the reviewer must decide or grant next.
Do not restate the review back at the reviewer, and do not send an acknowledgement between rounds — the next message is the candidate.
Pitfalls
- Editing the rejected patch in place, then losing the ability to show what the reviewer's bytes were.
- Fixing a nearby defect the review did not ask for; the extra hunk becomes the next rejection.
- Resubmitting a banned conclusion in softer words because the grep only matched the original phrasing.
- Presenting a stricter or newer measure as a refutation of an older attributed figure from a different definition, or silently reconciling the two.
- Turning a string-search zero into "the provision does not exist" instead of "not in this version".
- Recording a 403 as an enduring property of the source so later rounds inherit a false blocker.
- Reporting apply-check success as if content had been accepted.
Supporting files
scripts/candidate_diff.sh— buildvNpatches from pristine vs edited copies with stable labels, then print numstat and rungit apply --checkfrom the repo root.
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.

