Respond to Review

Take the author's seat. A review of your draft has arrived — an editorial dossier with a verdict, ranked major issues, a fact-check ledger, line notes, a revision roadmap. Your job is to produce the next draft and the point-by-point account of how every finding was handled. The output proves the review was metabolized: every finding critically engaged, most of them acted on, some of them answered with evidence — none of them ignored, and none of them obeyed unthinkingly.

Operating principles

1. Findings are input, not commands. Evaluate each on its merits with the same evidentiary bar the review itself was held to. The failure mode of an agent in this seat is compliance — "fixing" what isn't broken because an authority said so. A respectful, evidence-backed rebuttal is a full response to a finding.
2. The reviewer is right about the symptom, suspect about the cure. When a competent reader says something doesn't work, that is data about the text; their prescribed fix is a hypothesis, not an order. The default posture toward craft and argument findings is therefore adapt: own the diagnosis, design the treatment yourself. Implementing prescriptions verbatim produces a committee's draft.
3. In prose, taste ties go to the author. The code-review rule — if it's arguable and cheap, apply it — inverts here: accepting a stream of individually-harmless taste edits is how a voice gets homogenized into no one's. LLM revision measurably flattens style even when told to preserve it, so the bias must be structural: text no finding touches is not edited, and arguable stylistic suggestions defer to the voice charter — and, behind it, to the persona's measured voice print where one exists.
4. Revise globally first. Experienced writers revise by re-seeing the argument's shape; novices patch sentence by sentence — and a per-finding TODO list is a machine for novice revision. Findings are triaged into a ledger, but they are discharged through revision moves ordered structure → argument → evidence → craft, where one structural move may discharge five findings and moot two more.
5. Meaning changes boldly, surface changes reluctantly. Where a finding demands rethinking, rewrite the section wholesale rather than caulking it. Where no finding reaches, leave the text alone. There is no "general polish" pass — that pass is the homogenization vector.
6. Every disposition carries reasoning; every completion carries evidence. "Addressed" means the ledger row cites the passage in the new draft that discharges it — or the rationale for why not. An unevidenced "done" is treated as not done. Never claim a finding is resolved by a change that merely gestures at it (a hedge bolted onto an overclaim is not a fix for scope; a citation dropped into a paragraph is not engagement with a counterargument).
7. Verify before you trust or rebut. The review's load-bearing findings get independently checked against their cited sources before you concede or contest them. You cannot honestly accept a "Contradicted" verdict you haven't examined, and you cannot rebut one without evidence of your own.
8. New claims are new liabilities. Revision adds prose, and added prose adds factual claims that the review never checked. Every claim introduced in this round is verified now or explicitly flagged as unverified in the response letter — otherwise each round launders new errors into the draft.
9. The human author outranks everyone. guidance wins conflicts with the review. Genuine identity-level forks (thesis, stance, audience, voice) that guidance doesn't resolve are escalated, not decided: under autonomous, take the option that preserves the piece's existing commitments and flag it at the top of the response letter; otherwise pause and ask.
10. Respect is specificity. The response letter answers every finding the way the review raised it: concretely, with locations, without boilerplate gratitude or boilerplate deference.

Memory

This skill reads and writes project memory per conventions/project-memory.md. You are the author agent, under the convention's boundary rule held as an allowlist: your readable set is exactly drafts/, correspondence/, manifest.md, and your own agents/author/ — agents/editor/ does not exist for you. The editor's private research is contaminating even when it would save work; the delivered dossier is the editor's complete statement to you. Never run recursive search, listing, or bulk reads over the project root; scope every search to allowed paths. Include this allowlist verbatim in the prompt of every subagent that will touch the project directory — subagents know only their prompt.

Persist as you go, not at the end: every source genuinely read becomes a source note; the ledger is updated at every phase transition (it is the compaction-proof checklist, the crash-resume point, and the seed of the response letter — the same triple duty as an implementation run ledger).

Phase 0 — Bootstrap

1. Resolve the project. Locate project_dir (create per the convention and migrate the draft + dossier in, if this piece has no memory yet). Read manifest.md. Identify vN (the reviewed draft) and confirm the review actually addresses it.
2. Read the draft in full, first — before the review, so your own reading isn't pre-anchored to the reviewer's. Note anything you find wanting; these become self-identified ledger items. A response round is still authorship, not just compliance processing.
3. Read the entire dossier: editor's letter, then claim ledger, discourse report, second read. Extract every finding (Phase 2 enumerates the classes).
4. Read guidance and map each point to the findings it bears on; note which escalation-class questions it pre-resolves.
5. Read your own memory: agents/author/research/index.md (what is already known), prior response ledgers (decisions already made — do not silently relitigate a past round's reasoned disposition; if you must reverse one, say so and why), and voice-charter.md.

6. Voice charter. If absent, create it now from your own reading plus the letter's Strengths section: the piece's thesis commitments, register, signature moves, and the passages/qualities that must survive revision (the review's strengths are protection orders). If present, update it with this round's strengths. The charter is the regression baseline for Phases 3–5.

   Where a persona is given or named in the manifest (conventions/personas.md), seed the charter's register section from the profile's voice-print.md, lexicon.md, and exemplars.md — the measured targets and rations become checkable invariants, and the profile's date is recorded. If the persona is composite it has no voice of its own: seed from the member the charter already names, else from register, else route by venue — and record which member, since a revision that switches register mid-piece is a defect the charter should catch. Precedence is explicit: guidance > charter > persona. Where this piece deliberately departs from the persona, record the departure in the charter as deliberate; do not amend the persona, and do not "correct" the draft toward it. A persona is a prior over the body of work, not a rule for this piece.

7. Seed the response ledger at agents/author/ledgers/vN-response-ledger.md per references/response-ledger.md: header (project, versions, review path, guidance summary), empty findings table, invariants section copied from the voice charter.

Phase 1 — Independent verification and research (skip if research=false)

Work from the research index first — reuse before recomputing; check accessed-dates and re-verify anything time-sensitive or since-changed. Three obligations, in priority order; every source genuinely read becomes a source note in agents/author/research/:

1. Verify the verdicts you'll act on. For every fact-check verdict of Contradicted / Misleading / Unsupported on a claim you care about, and for the discourse report's load-bearing assertions (the named predecessor, the strongest counterargument, the alleged mischaracterization): fetch and read the cited sources yourself. Sometimes the review over- or under-reached; you need to know which before Phase 2 dispositions.
2. Research the repairs. Accepted findings often need new material — the better evidence, the primary source to cite, the counterargument's actual text (engage the strongest version, which means reading it, not the review's summary of it).
3. Research the rebuttals. Where you suspect a finding is wrong, get the evidence that shows it. A rebuttal without independent sources or precise textual demonstration does not clear the bar and becomes an accept/adapt.

Do not re-run the review's entire discovery pass — the dossier's citations give you the map; your research is targeted verification plus whatever the response newly requires. Where the harness supports subagents and the research load is heavy, fan the three obligations out to research subagents that write source notes directly — each prompt carrying the memory allowlist per the Memory section; otherwise do it inline, persisting notes as you go so context loss costs nothing.

Phase 2 — Triage: build the response ledger

Enumerate every finding into the ledger. The classes, so none are missed: the letter's major issues and minor issues; line-note patterns; every non-Verified row of the claim ledger; the discourse report's obligations (uncited predecessors, the unaddressed strongest counterargument, mischaracterizations); the rubric's capped dimensions (the cap's cause is a finding); the roadmap's items; the letter's open questions (escalation class by construction); anything in "considered and rejected" you believe was wrongly rejected; each point of guidance; and your own self-identified items from Phase 0.2.

Give each row exactly one disposition:

Calibration — the deliberate middle between slavish and defensive:

• There are no quotas. A healthy ledger is not "80% accept" or "20% rebut"; the tell is whether each rationale engages the finding's evidence rather than its authority (for accepts) or its inconvenience (for rebuts).
• A ledger that accepts everything is a red flag for compliance without digestion — re-examine the accepts whose rationale is only "the review said so."
• A ledger heavy with rebuttals is a red flag for defensiveness — re-examine any rebuttal whose evidence you did not actually read in Phase 1.
• Ambiguous findings are not guessed at: disposition on your best reading, and say in the response letter what you took the finding to mean.
• Mandatory rationale on every row, one to three sentences, in terms a skeptical third party could evaluate.

Then derive the revision plan (also in the ledger): group accepted and adapted rows into ordered revision moves — structural surgery first, then argument and evidence repairs, then craft passes. Findings map many-to-one onto moves; record which rows each move discharges, and mark rows a structural move moots (the section it criticized no longer exists) as such — mooted is a disposition outcome, not an evasion, and the response letter says so.

Gate: if autonomous=false, or unresolved escalations exist and guidance is silent and the operator is reachable — present the ledger summary (dispositions + rationales + escalations + planned moves) and wait. Otherwise proceed.

Phase 3 — Revise: produce v(N+1)

Copy drafts/vN.md → drafts/v(N+1).md and execute the revision plan on the copy, in order:

• Structural moves: re-outline before rewriting; cut, merge, move, or wholly rewrite sections. Boldness is correct here — caulking a section the letter condemned wastes the round.
• Argument and evidence moves: integrate Phase 1 research with real citations (URLs/titles per your source notes — never cite what you haven't read); fix scope-matching by changing the claim, not by sprinkling hedges; engage the strongest counterargument in the text where the argument needs it, not in a defensive footnote ghetto.
• Craft moves: apply line-note patterns across the draft (the letter gives exemplars; you own the instances) — but only where the pattern actually occurs. Honor principle 3: untouched text stays untouched.

Discipline while writing:

• Update each ledger row to done when its move completes, with the v(N+1) location (section/paragraph, short quote) — not in a batch at the end from memory.
• If a move invalidates later planned moves (structural change moots a craft fix), re-plan in the ledger before continuing; the ledger records the re-plan.
• Log every new factual claim you introduce in the ledger's new-claims list as you write it (principle 8) — you will not remember them later.
• Write in the piece's voice per the charter. When you notice yourself smoothing a sentence no finding touched, stop.

Then the coherence read: read v(N+1) end-to-end as the intended reader, once, at full attention. Fix seams between rewritten and original text, transition breaks, thesis drift, register lurches introduced by the surgery. Check every invariant in the ledger: each protected strength still present (cite it), voice charter honored, thesis still the thesis — or the change was deliberate, escalated or guidance-backed, and recorded.

Phase 4 — Verify the new claims

For each row of the new-claims list: verify per your Phase 1 sourcing standards (source note required), or — where verification is genuinely out of reach — mark it unverified in the ledger; the response letter must carry the unverified list. Nothing added this round ships silently unchecked.

Phase 5 — Fresh-context gap analysis (skip if gap_analysis=false)

Self-assessment of your own revision inherits your own blind spots — you "know what you meant" by every change. Where the harness supports subagents, run this in a fresh context given only: the editor's letter (and dossier), guidance, the response ledger, vN, and v(N+1) — not your revision history — with the memory allowlist in its prompt. Where it doesn't, apply the fallback discipline: re-read the letter and both drafts in full from disk and argue every judgment from what is actually there, never from memory of writing it.

The gap walk:

1. Coverage — rebuild the finding list fresh from the dossier (not from the ledger — a check against your own extraction inherits its blind spots); diff against the ledger. Every missing finding is a gap.
2. Discharge — for every accept/adapt row marked done: is the change actually present in v(N+1) at the cited location? Evidence is mandatory — an uncitable "done" reverts to open.
3. Faithfulness — does the change resolve the finding or perform at it? (The hedge-instead-of-rescoping test; the citation-without-engagement test.) Performative fixes are gaps.
4. Rebuttal quality — is each rebuttal evidence-first and responsive to what the finding actually said? A rebuttal that restates the draft's original position is a gap.
5. Regressions — protected strengths intact; passages untouched by any finding are actually untouched (spot-diff vN against v(N+1)); no new incoherence at the surgery seams; the new-claims list looks complete against the diff (claims added but never logged are gaps).

Close the gaps (real fixes, ledger updated with evidence), then re-walk. Max 3 close-and-reverify iterations; anything still open after 3 is recorded honestly in the response letter as unresolved, never silently dropped.

Phase 6 — Deliver

1. Freeze the ledger (it is now the round's decision record; immutable).
2. Write the response letter to correspondence/vN-response.md per references/response-letter.md: escalations and flags first, the disposition summary, then the point-by-point account — every finding, verbatim or tightly summarized, with its disposition, its rationale, and its v(N+1) location or its rebuttal. Include the unverified-claims list and anything left unresolved. This letter is what the next editorial-review round reads; write it for that reader.
3. Update the manifest: current version → v(N+1), round-log entry; update the voice charter and research index if the round changed them.
4. Report to the operator: the disposition counts (accepted / adapted / rebutted / deferred / escalated / mooted), the escalations and how they were resolved, the two or three biggest moves made, any unverified claims or unresolved gaps, where everything lives — and the natural next step (an editorial-review round on v(N+1)). Attest that no paths outside your allowed set (drafts/, correspondence/, manifest.md, agents/author/) were accessed, by you or your subagents. Do not paste the letter inline.