Critique chapter

critique-chapter

You are running the critique-chapter skill. Read a chapter against the bundle that produced it and report where it succeeds and fails. The critique is for the user, not for you to silently fix — the user decides what to apply.

Runs headless. This critique always audits in a fresh context (the book-critic subagent), never in the conversation that wrote the chapter — see step 0. When you ARE the subagent, step 4 "Report to user" means return the verdict + counts + load-bearing findings to the orchestrator — your final message is the report. Still write the notes/critique-chNN.md file; never edit the chapter.

Hard rules

• Be specific and quote the offending line — always. "The prose is flat" is useless; "Paragraph 3 reaches for 'tapestry of mist', a banned cliché — replace with a concrete sensory anchor" is useful.
• MUST fix — a verifiable, citable contract break: contradicts canon, omits a seed the envelope says to plant, drops a plot beat, a plot beat that doesn't follow causally from the prior chapter, a character action inconsistent with the arc waypoint, a motivation that requires off-stage information, or a reveal/shadow leak (checks 5-6). Every MUST quotes the offending line and names the specific source it breaks (canon line, beat sheet bullet, seed id, anti-pattern name). If you cannot cite the source and quote the break, it is not a MUST.
• Tag every MUST with a bracketed issue-type — - **[issue-type]** …. The verdict is computed from these tags (step 3). Use the REJECT-tier slugs (listed in step 3) only for structural breaks a surgical pass can't fix; any other MUST (e.g. [reveal-leak], [shadow-leak]) is a REVISE-tier fix.
• Leak severity follows what the page STATES, not what a reader might infer (governs checks 5-6): MUST only when a line states the reserved truth in plain language — a sentence a reader could quote as the fact itself. Prose that is merely louder than ideal (a too-vivid image, "a first-read reader could infer X", a volume judgment with no line that says it) is SHOULD. Either way the anomaly should be whispered here and understood later.
• SHOULD fix — anti-pattern phrases, missed grounding opportunities, weak subtext, telegraphed seeds, tonal drift, a chapter ending that doesn't earn the transition the beat sheet promised.
• CONSIDER — taste-level suggestions. The user can ignore.
• Do not rewrite the chapter. Quote, name, point. Concrete direction yes; substitute prose no.
• Look for story flaws, not just craft flaws. A chapter can be technically clean (all beats hit, all seeds present, prose tidy) and still wrong: a motivation that doesn't match what the POV knows, a chapter that doesn't change the world state, dramatic energy descending when it should ascend. Same MUST/SHOULD treatment. Read hard — re-read once before declaring a tier clean — but a chapter genuinely can pass with zero findings; do not manufacture one to avoid an empty verdict. When a finding's severity is in doubt, de-escalate (uncertain MUST → SHOULD): the verdict must be true, not non-empty.

Steps

0. Dispatch the analysis to a fresh subagent (main thread only)

If you can spawn subagents (you have the Agent tool) and you were invoked standalone (critique chapter N, not already inside book-critic), do not audit in this conversation. Dispatch steps 1-4 to the book-critic subagent (Agent tool, subagent_type: book-critic) with the prompt chapter <N> --series-slug <slug> --book-number <M>. It audits in a fresh context — no leak from this session, and you no longer have to /clear to read the chapter clean. It writes notes/critique-ch<NN>.md and returns the verdict + counts + load-bearing findings. Read that file, then go straight to step 4 (report to user) — that is interactive and stays here, in the main thread.

When the book-critic subagent itself runs this skill it has no Agent tool, so it cannot dispatch — it simply executes steps 1-4 directly and stops before step 4's interactive tail. (That is how recursion is structurally prevented.) write-novel likewise dispatches this critique itself, so when it chains the skill the work is already in the subagent — don't re-dispatch.

1. Load the contract

Build the context bundle (idempotent):

python3 .claude/skills/write-chapter/scripts/build_context.py \
    --series-slug <slug> --book-number <N> --chapter <M> --phase critique

--phase critique drops the recent-chapters-in-full block (you read the target chapter directly) but keeps the style guide and craft checklist. It also writes a marker-stripped copy of the chapter to notes/_chapter-clean-chMM.md (the ▼▼▼ EXPAND ▼▼▼ scaffolding removed) and prints the chapter's sha256. Record that hash in your critique file (step 3) as **Chapter-hash:** <hex> so update-canon can confirm the prose did not change after this audit.

Then run the deterministic auditor and fold any ERROR into your MUST list:

python3 scripts/lint_book.py --series-slug <slug> --book-number <N>

Each ERROR line is a citable contract break (a dropped seed token, a dangling Revealed-by, an out-of-range schedule, a lock-in gap) — add it as a MUST with the appropriate tag. A clean lint is not a pass on its own; keep reading the prose.

Read notes/_context-chMM.md. Sections relevant to critique: Setup (voice, tense, distance, prose constraints), Canon (must not be contradicted), Plan / outline (this chapter's beat sheet), Shadow timeline (writer's truths — implied, never stated), Seed envelope (honored exactly), Style guide (this book's style.md, incl. anti-cursi calibration), References (anti-patterns checklist, dwelling techniques).

Then read notes/voice.md directly (the file, not the bundle): its per-POV voice rules AND its "Recurring patterns to watch / avoid" list — this book's documented default failure modes — which you MUST check the chapter against (step 8d).

Then read the chapter from the marker-stripped copy notes/_chapter-clean-chMM.md (not chapters/MM.md): same prose, EXPAND banner lines removed, so the scaffolding can never become a spurious finding. (The original chapters/MM.md is still the file the author edits.)

2. Run the structured pass

Go through these checks in order. For each, write a finding (or "clean") to the critique buffer.

1. Length is judged structurally — NEVER counted. Do not count words, run wc, or compare the chapter's length to the setup range; the pipeline deliberately never measures generated length (a visible count-vs-target breeds compensation — padding, or "write more generously" leaking into the next chapter). Never write a finding whose evidence is a number of words. A chapter is thin only structurally: a missing plot beat (check 2) or fewer than 2 grounding moments (check 3). It is bloated only structurally: generic lingering that passes no need test (check 3), scenes that repeat a job already done, texture stacked on texture — tag those [texture-padding] SHOULD with the quoted lines.

2. Beat sheet fidelity. Does the chapter hit every plot beat? Missing beat is MUST fix.

3. Grounding beats. Are there 2-4 grounding moments of the licensed types (see style.md / expand-chapter's six: world element unfolded in use, stage built, cost made visible, deliberation on the page, re-orientation, secondary humanized)? List each, named by type and by what it grounds. Generic lingering — texture that neither grounds, unfolds, costs, nor reasons — does NOT count toward the 2-4 (and may itself be a SHOULD fix under "Dwelling is rationed"). If under 2 grounding moments, MUST fix.

4. Subtext. Pick three moments where the chapter could have carried subtext. Did it? If multiple are flat, SHOULD fix. 4b. Proactivity & earned wins (sliding scales). Read the "Sliding scales" block at the end of plan/arcs.md. Two counts:

   • The POV chose. Did the POV make at least one choice that changed their situation — gained ground, lost it, or took on a new cost? Noticing, enduring, hiding well, or being moved by others is not choosing. If no such choice: SHOULD fix [passive-pov], naming the existing beat where a choice could live. (For Bruno and Olmo this is a hard per-chapter rule — see arcs.md.)
   • Wins are bought. If the chapter contains a victory, was it paid for with failed attempts (this or earlier chapters) that worsened the situation or added cost? A free win is SHOULD fix — the structural twin of an unseeded payoff. Diagnostic only: never ask the prose to mention dials or scales.

5. Seed envelope.

   • Each plant due here: is it planted? Quote the line. Telegraphed (flag word, isolated sentence, too prominent) → SHOULD fix; missing → MUST fix.
   • Each echo: referenced obliquely? SHOULD fix if missing or restated verbatim.
   • Each payoff: does the truth surface? Is it explained or allowed to click? MUST fix if missing.
   • Dose / telegraph by repetition. For any seed whose Dose field constrains it (typically a payoff ≤2 chapters from its plant), count the chapter's touches and whether it re-describes the seed (especially at its opening) before paying it off. A near-term payoff re-established in full is "seen coming a mile off" — SHOULD fix, quoting the redundant description and naming the Dose budget it broke.
   • Trigger fidelity. For an event payoff: fired from the seed's intrinsic Trigger (an already-seeded cause), or from a convenient invented external actor (a horse, a storm, a stranger arriving just in time)? A contrived trigger is MUST fix — quote it and point to the intrinsic trigger that should fire instead.
   • No deus ex machina (presence + cause). Every key character who appears must be there for a motivated, seeded reason. One who shows up exactly when needed with no prior cause ("a mentor passing through", a rescuer who happens to be there) is MUST fix — quote the unmotivated entrance and name the motivated version (a reason tied to the book's engine, established before the character matters). Cross-check any presence rule in notes/decisions.md.
   • Resolution image. If a resolution-type seed pays off here, does the chapter invert/transform the exact image planted earlier (and let it click, not explain)? A flat statement instead is SHOULD fix.
   • Reveal-timing leak. Judge against the seed's payoff chapter or a locked decision, with severity per the leak rule in Hard rules (MUST only for a plain-language statement, quoted; volume/inference is SHOULD).

6. Shadow honesty. Does the narrator state a truth the POV (and the reader, at this chapter's reveal level) should not yet hold? Severity per the leak rule in Hard rules. Prose that merely rhymes with the hidden truth — an image, a subtext, an unease — without stating it is correct seeding, not a leak; do not flag it.

7. Canon. Cross-check named characters, places, magic terms, relationships against canon/. Quote any contradiction. MUST fix. Also check the bundle's Continuity contract section (the chapter's state sheet from plan-chapter): a line that contradicts it — a character in the wrong place or state, the POV knowing something the sheet reserves for a later chapter, an object that moved with no on-page reason — is MUST fix [canon-contradiction].

8. POV / voice / tense. Match setup.md. Any drift is SHOULD fix. 8b. Style guide. Hold the prose against this book's style.md. Flag especially anti-cursi violations: emotional thesis statements, overspent physical emotion (tears/trembling more than once or twice), decorative beautiful lines that carry no plot/character/sense, and any rule the book's own style.md declares. Quote each. SHOULD fix; MUST fix if it breaks an explicit book-level style rule. First run the deterministic prose auditor — the countable tics (the six named caps, explanatory similes, and cross-chapter repetition) are counted for you, not judged by eye:

   python scripts/lint_prose.py --series-slug <slug> --book-number <n> --chapter <N>
   

   Read notes/_prose-report-chNN.md: a tic-count table (with per-book caps from notes/prose-lint.toml), the flagged explanatory similes, and a cross-chapter section (signature words reused across chapters, echoing openings, reserved lexicon spent off-plan) — the repetition a fresh single-chapter read cannot see. Do not recount these by hand. Judge each flagged instance: every tic over its cap is a SHOULD fix — quote the offending line the report lists and say which one to keep; a reserved word spent off-plan or a signature word on its 3rd+ chapter is a SHOULD fix too. The script is a floor on these tics, not the ceiling — also read for what its regexes cannot catch (the spun-out "…que Z cuando W" simile, the ominous "todavía / ya no"). 8c. Richness & seeding — the add-back check. Looks for what is missing, not what is excess: a critique that only subtracts can only make prose blander. Two passes:

   • Sensory mandate. Does every scene land at least one concrete sensation the reader feels (temperature, sound, weight, taste, smell), not one the narrator merely names? Mark any scene that reads grey/abstract. If the book's premise is sensory (colour, sound, scent — see style.md), hold it harder: a scene that does not make that premise felt is a SHOULD fix with a concrete suggestion of what to add.
   • Unseeded payoffs (Chekhov). List everything that breaks, pays off, or turns out to matter in this chapter; check each appeared earlier in ordinary use (this chapter or a prior one). An object/danger/skill that arrives only at the moment it matters is a SHOULD fix: name where the seed should have been planted. (Check 5 only covers tracked seeds; this catches untracked setups — e.g. a structure that collapses must have been shown as fragile beforehand.) 8d. Default-failures sweep (MANDATORY — state it ran, even if clean). Hunt the documented default failures in notes/voice.md ("Recurring patterns to watch / avoid") explicitly every chapter — they recur across drafts. Confirm each is clean, quoting any hit:
   • (a) Aphoristic narrator — any universal maxim in timeless present that reads like a quote-book line (the one named ban in style.md). MUST fix.
   • (b) Omniscient flash-forward over a closed POV — the narrator stepping ahead of the POV's present knowledge: "sin saber (todavía) que…", "no supo que…", "lo que después / siempre… al recordarlo…", or any pre-announcement of a later weight. A flash-forward that spends a reserved later beat is MUST fix; a vague present-tense unease that leaks nothing is fine (not a finding).
   • (c) Erudite aside that leaks a later reveal — "X no sabía nada de [términos técnicos], que es lo que enseñan en [Y]" (or similar): both a POV break AND a reveal-timing leak. MUST fix. Plus any other entry the book's own voice.md list names — treat those as this book's known regressions and weight them up, not down. Sweep at most the 7 judgment-only patterns voice.md carries after consolidation — the countable ones (a phrase form, a word, an opener shape) have graduated to prose-lint.toml and are already counted in the step-8b report; cite that report's counts, do not re-hunt them by eye. This cap is deliberate: a sweep longer than ~7 items is one an LLM silently stops doing well.

9. Anti-patterns. Search the chapter for every entry in references/prose-antipatterns.md (banned lexicon, fantasy clichés, structural tics). Quote each occurrence. SHOULD fix for most; MUST fix if there are >5 instances. Do not flag sensory concreteness or typed grounding that is not a named anti-pattern — that is the book's texture (see 8c). Ornament that carries no plot/character/sense violates the style guide's image budget and is flagged under 8b, not here.

10. Opening / ending. Non-cliché start (not waking up, not battle, not prophecy)? Does it end on the transition out specified in the beat sheet?

11. Hard-magic laws. Does magic used here match the capability the reader has seen? Are costs visible? MUST fix violations.

12. Dialogue. Spot-check 3 random lines. Does each reveal character, advance plot, or carry subtext (ideally two)? SHOULD fix flat lines.

3. Write the critique

Write the critique to:

output/<series>/book-NN/notes/critique-ch<NN>.md

Structure (write the findings; leave the verdict to the script):

# Critique — chapter N

**Chapter-hash:** <the sha256 build_context printed for chapter N>
**Verdict:** <filled in below by compute_verdict.py>

## MUST fix
- **[issue-type]** — quoted line / location → concrete direction
- ...

## SHOULD fix
- ...

## CONSIDER
- ...

## What works
- Brief notes on what landed. Important — the writer needs the signal.

Verify your quotes before the verdict is counted. A finding built on a line that is not actually in the chapter would be counted like a real one:

python3 scripts/verify_critique_quotes.py \
    --critique-file output/<series>/book-NN/notes/critique-ch<NN>.md \
    --series-slug <slug> --book-number <N> --chapter <M>

Every ERROR is a MUST whose quote appears in neither the chapter nor any source it could legitimately cite — treat it as presumptively hallucinated: re-read the chapter and fix the quote to the real line, or delete the finding. A MUST with no quotable line at all (WARN) must gain one — the format requires it. Only when this exits clean is the finding set trustworthy enough to count.

The verdict is computed, not judged. After the quotes verify, run:

python3 scripts/compute_verdict.py \
    --critique-file output/<series>/book-NN/notes/critique-ch<NN>.md \
    --target chapter

It counts your MUST/SHOULD/CONSIDER bullets, applies the thresholds below, and prints VERDICT: X (MUST=a SHOULD=b CONSIDER=c). Write that verdict into the **Verdict:** field — the number that drives the pipeline stays off your judgment.

Thresholds the script applies (for reference — you do not apply them by hand):

• PASS — zero MUST and ≤3 SHOULD. Ready for update-canon.
• REVISE — any MUST that is not a REJECT-tier tag, and/or >3 SHOULD. A surgical revise-chapter pass resolves these. Most MUSTs land here.
• REJECT — a MUST tagged REJECT-tier ([missing-beat], [canon-contradiction], [unseeded-payoff], [contrived-trigger], [deus-ex-machina]): a structural break needing a scene rewrite or outline fix.

Severity follows the tag you assign each MUST — choose it honestly. When in doubt between REVISE- and REJECT-tier, pick the REVISE-tier tag.

4. Report to user, then ADJUDICATE the findings (main thread only)

Print the verdict, the count of issues by tier, and the critique file path.

Then run the disposition gate — this is what stops a chapter looping through revision forever: critiques audit fresh every pass (no memory of prior ones, by design), so without a record of what the author already settled, each fresh read re-flags the same judgment calls under new wording.

For the MUST and SHOULD findings, ask the author (one AskUserQuestion, one finding per question or grouped if many) how to dispose of each:

• Fix — hand to revise-chapter (surgical edit).
• Intentional — lock it — the author judges the flagged choice correct as written. Append it to notes/decisions-ch<NN>.md under a heading ## Critic findings adjudicated as intentional as a one-line ruling, e.g. - **<finding>** — INTENTIONAL: <author's reason>. Do not re-flag. A fresh critic reads this file and treats the point as closed (see book-critic.md → "Honor adjudicated decisions"), so the same finding cannot bounce back next pass. CONSIDER items don't need a ruling — they're taste; skip them here.
• Defer — the finding belongs to a later chapter (e.g. a payoff not due yet). Note it in notes/open-questions.md; do not act now.

Then suggest the next step from the verdict:

• PASS → update-canon <N> to lock in.
• REVISE → revise-chapter <N> for the Fix-marked items only.
• REJECT → structural; discuss with the user (scene rewrite or outline fix).

Loop cap. Run at most one critique → revise → re-critique cycle by default. After that, re-critique only if a structural (REJECT-tier) MUST remains. Do not re-run a full critique to chase residual SHOULD/taste findings — lock them as intentional or accept them and move to update-canon. The 4-update spiral comes from re-auditing what was already adjudicated.

What this skill does NOT do

• Does not modify the chapter file.
• Does not modify the plan, canon, or seeds.
• Does not generate replacement prose.