Critique plan

critique-plan

You are running the critique-plan skill. The plan is finished (or the user thinks it is). Your job is to find what's wrong with it before any prose is written.

The user has explicitly asked for this to be adversarial, not validating. Push back. The plan exists to be stress-tested.

When to invoke

• "Audita el planteamiento" / "critique the plan" / "is this plan ready?" / "review the book plan".
• After plan-book and before write-chapter 1 — even on user request.
• Also reasonable as a re-check after a major plan revision.

Hard rules

• Be specific. "Structure is weak" is useless. "Act 1 spans 9 chapters but no character arc shows movement until chapter 7" is useful.
• Quote the offending line / cite the chapter number / name the seed id. Always.
• Tier ruthlessly:
  • MUST fix — breaks the contract: missing climax decision, seeds with no payoff, principals with no decision moment, magic with no cost or limits, outline chapters with TODO-only beats.
  • SHOULD fix — weakens the book: thin texture-beat plans, midpoint that doesn't overturn anything, faction declared but never used, action stacking, telegraphed seeds.
  • CONSIDER — taste-level. The user can ignore.
• Do not rewrite the plan. Quote, name, point. Direction yes; substitute prose no.
• Adversarial bias. When unsure, escalate the tier. A plan that passes critique-plan with zero issues is suspicious; you missed something.
• No leak from prior reviews — audit FRESH every time. Do not read the previous notes/critique-plan.md or any earlier _audit-plan.md before forming your own findings. Each pass must be independent: a fix the author applied between runs has to survive a clean re-read, not a memory of "we already decided this is fine". Anchoring on the last verdict is an information leak that hides regressions. The only audit you trust is the one you generate this run (step 1). Overwrite, don't merge.

Steps

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

If you can spawn subagents (you have the Agent tool), do not audit in this conversation. Dispatch steps 1-4 to the book-critic subagent (Agent tool, subagent_type: book-critic) with the prompt plan --series-slug <slug> --book-number <N>. It audits in a fresh context — no leak from this session, and you no longer have to /clear to re-audit in clean. It writes notes/critique-plan.md and returns the verdict + counts + load-bearing findings. Read that file, then go straight to step 5 (report) and step 6 (the AskUserQuestion solution menu) — those are interactive and stay 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 5. (That is how recursion is structurally prevented.)

1. Run the deterministic audit

python3 .claude/skills/critique-plan/scripts/audit_plan.py \
    --series-slug <slug> --book-number <N> \
    --output output/<series>/book-NN/notes/_audit-plan.md

The audit checks mechanical truths: gating fields, TODO-only chapters, seed validity, decision-chapter presence on each arc, canon entries absent from outline, magic completeness. Read this file — it gives you the baseline list of MUST-fix items.

2. Read the plan in full

Read in this order:

1. setup.md
2. plan/outline.md
3. plan/shadow.md
4. plan/seeds.md
5. plan/arcs.md
6. canon/characters.md, canon/factions.md, canon/magic.md, canon/world.md, canon/timeline.md
7. References for what "good" looks like:
   • references/fantasy-beats.md
   • references/seed-craft.md
   • references/magic-design-checklist.md
   • references/hard-magic-laws.md

3. Run the qualitative pass

Go through each of these categories. For every check, produce a finding (or "clean") for the critique buffer.

3a. Structure

• Act boundaries — does act 1 take enough time to inhabit the world (slow-immersion), or does it rush?
• Midpoint — does something the reader believed get overturned at the named midpoint chapter? Quote the chapter's beat sheet to prove it.
• All-is-lost — is there a named chapter where everything appears beaten? Or does the plan skip from rising action to climax?
• Climax — does the climax decision rest on something that was planted earlier (a magic capability seen, a relationship built, a seed paid)? If the climax depends on a tool that appears in its own chapter for the first time, that is MUST fix.

3b. Plot / trama

• Causal chain — does each chapter's plot beat follow from the prior chapter's outcome, or are there leaps?
• Motivation — at every plot beat, can the POV character plausibly want this with what they know at that point? Flag any decision that requires offstage information.
• Coincidence load — count the moments where chance moves the plot. More than 2 in act 2 is a SHOULD fix.
• Antagonist competence — does the antagonist have a coherent plan in the shadow file? If shadow says "they happen to be there", it is MUST fix.

3c. Ritmo / pacing

• Chapter function — every chapter must do something for the larger arc. Flag chapters whose function is "more of the same".
• Action stacking — two action-heavy chapters back-to-back without a quiet chapter between is SHOULD fix per references/fantasy-beats.md pacing rules.
• Texture distribution — count chapters with explicit texture beats planned. If many lack texture beats, the book will come out short and dry. SHOULD fix per chapter.
• Dwelling in act 1 — at least 60% of act 1 chapters should plan a texture beat. If under, the book won't feel inhabited.

3d. Magic system

• Source, mechanic, costs (≥2), hard limits (≥3), thematic question, three escalation tiers — every one of these must be present and specific. MUST fix any missing.
• Costs visible in prose — does the outline name which chapters show the cost in action? If costs are stated in canon but never appear in outline texture beats, the magic will feel cheap.
• Apex tier visible only at climax — flag if apex magic appears in act 1 or 2. The climax loses its ceiling.
• Thematic question forced at climax — does the climax beat sheet put the thematic question in the protagonist's hands? MUST fix if not.

3e. Characters

• Distinct arcs — read each principal's want/need/lie/decision. If two principals have the same lie or the same decision shape, the book has only one arc. SHOULD fix.
• Decision chapter alignment — the protagonist's decision moment should sit at the climax. Secondary characters' decisions should feed into the climax. Flag misalignments.
• POV economy — is every declared POV needed? A POV with fewer than 3 chapters of presence is CONSIDER. A POV character who never carries a plot hinge is SHOULD fix.
• Physical specifics — three concrete details per principal, no clichés. Vague entries are SHOULD fix.

3f. Subplots

• Count subplots. If 0, flag — epic fantasy needs ≥1 subplot to breathe outside the protagonist.
• Touchpoints — each subplot must touch the main plot at ≥3 points. Quote them. MUST fix if a subplot is hermetic.
• Resolution timing — subplots should resolve before or during the climax. Any that resolve after are SHOULD fix (energy drains).
• Distinct theme — subplot themes must differ from the main theme. If both arcs argue the same thing, the subplot is decoration.

3g. Foreshadowing / seeds

• Grimoire §14 coverage — the audit cross-references the grimoire's §14 loaded-gun table against the plan: every loaded gun whose "Siembra en" includes this book MUST have a seed tagged **Obligatory:** §14 <name>. Any the audit lists as missing is a MUST fix — a worldbuilding promise the plan forgot to plant (this is the count check that catches a dropped seed). A seed tagging a non-existent §14 row is a typo to fix.
• Series / trilogy coverage — for a book that is NOT the last of a series, the audit checks it actually seeds the books that come after it, not just resolves itself. It flags later books the plan plants nothing for, thin seeding into the immediate next book (the middle book needs its payoffs planted now), and a seed/truth floor raised by the cross-book burden. These are SHOULD fix — a trilogy opener that only resolves itself leaves the next book to inherit nothing. (None of these are fixed counts; they scale with chapters and number of later books.)
• Density — ≥ 8 seeds for an epic fantasy. Fewer is SHOULD fix.
• Distribution — plants weighted toward act 1 / early act 2; payoffs weighted toward act 2B / act 3. The audit shows per-act counts. Flag inversions.
• Telegraphing risk — read each seed's how_to_plant. If it sounds like the seed is the scene's purpose ("a strange old man with glowing eyes appears at the door"), it telegraphs. SHOULD fix, citing references/seed-craft.md.
• Echo gaps — if a seed plants in chapter 3 and pays off in chapter 20 with no echoes between, the reader will have forgotten it. SHOULD fix.
• Climax-as-payoff — does the climax pay off at least one major seed? If not, the climax floats. MUST fix.
• Trigger soundness (event payoffs). For every seed whose payoff is an event (something breaks, fires, collapses, is discovered), is there a Trigger field naming an intrinsic, already-seeded cause? A payoff with no trigger, or a trigger that is a convenient external actor (a horse, a storm, a stranger who only shows up to cause it), is SHOULD fix — the writer will improvise a contrivance at writing time. Cite the seed id and the missing/weak trigger.
• Telegraph by dose & distance. Flag any seed whose payoff is ≤2 chapters from its plant and that also carries echoes or a heavy how_to_plant — it will be "seen coming a mile off". Each such seed needs a Dose field constraining it to one quiet touch. SHOULD fix, citing the plant/payoff chapters.
• Emotional spine present. Are there ≥2 resolution-type seeds (a Resolution image: a planted felt-image the payoff inverts/transforms), so the book resolves and not merely concludes? Zero is SHOULD fix — the plot will pay off but the book will feel mechanical, "not woven".

3h. Worldbuilding economy

• Named characters who never appear in the outline — the audit flags these. Each is SHOULD fix (either use them or cut them).
• Named factions never used in outline — same.
• Named places never used in outline — same.
• Historical events declared but not echoed — historical weight exists to reverberate. If no chapter touches a declared past event, it is decoration. CONSIDER.

3i. Shadow honesty

• Overview — does the shadow overview actually reveal a different story than the outline? If shadow restates the outline in code words, there is no hidden layer. MUST fix.
• Master truths — ## SHADOW-TRUTH records, and enough of them: not a thin handful but coverage of the protagonist's hidden nature, each antagonist's agenda, each institution's real function, and each major subplot (~12-20 for a 25-chapter book; the audit flags thin coverage). Each truth needs a reveal path — either Revealed-by: carrier seeds (the schedule lives there, never re-scheduled in the truth) or a manual Confirm in: for exposition-only truths. A truth with neither can never reach the reader: MUST fix. Carrier seed ids must exist in seeds.md (audit flags unknowns). Reveal cap must be honest — a truth that pays off in a later book caps below confirmed.
• Grimoire §14b coverage — the audit cross-references the grimoire's §14b master-mystery table: every mystery introduced in this book MUST have a shadow truth tagged **Mystery:** <name>. Any the audit lists as missing is a MUST fix (a promised mystery the plan won't carry); a truth tagging a non-existent mystery is a typo.
• Reveal intensity — the cap names the reader's interior state (hidden → sensed → suspected → confirmed), never the writer's loudness. Flag any plan note that tells the writer to state a truth plainly ("hard hint", "make it obvious") — that defeats the shadow.
• Antagonist knowledge — does the shadow describe what the antagonist knows that the protagonist doesn't? SHOULD fix if silent.

3j. Consistency

• Cross-check names: do the principals in arcs.md match the names in canon/characters.md? Mismatches = MUST fix.
• Cross-check seed plant/payoff chapters against outline: does the named chapter exist and have a relevant subtext beat? Mismatches = SHOULD fix.

4. Write the critique

Write to:

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

Overwrite the previous file — never merge (see Hard rules: audit fresh).

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

# Plan critique — <book title>

**Verdict:** <filled in below by compute_verdict.py>

**Summary:** one-paragraph diagnosis.

## MUST fix
- **[issue-type]** — concrete finding with a quoted line or chapter
  pointer → suggested direction.
- ...

## SHOULD fix
- ...

## CONSIDER
- ...

## What works
- 3-6 brief notes on what the plan does well. Important — the user
  needs the signal too.

Tag each MUST with a bracketed issue-type. Use these REJECT-tier slugs for structural breaks no surgical edit fixes: [climax-unplanted] (climax decision rests on a tool first appearing in its own chapter), [shadow-single-layer], [principal-no-decision], [subplot-hermetic], [magic-missing-pillar]. Any other MUST (a missing Trigger to add, a seed to retag, a decision-chapter to realign) is REVISE-tier.

The verdict is computed, not judged. After writing the findings, run:

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

Write the printed verdict into **Verdict:**. Thresholds it applies: PASS = zero MUST and ≤6 SHOULD; REJECT = any REJECT-tier MUST; REVISE otherwise. When in doubt between REVISE- and REJECT-tier, pick the REVISE-tier tag.

5. Report to user

Print:

• Verdict and counts by tier.
• The 3-5 most load-bearing findings (one line each).
• Path to the critique file.

Then go straight into step 6 — do not just leave the verdict and wait. The contract is: critique-plan ALWAYS ends by offering possible solutions as an AskUserQuestion menu, not as free prose the author has to parse.

6. Offer solutions as an AskUserQuestion menu

After reporting, surface the findings the author must act on as choices with concrete solutions — this is the required closing move of the skill.

• PASS, zero MUST/SHOULD → one confirm question: "Plan looks ready — proceed to chapter 1?" with options write chapter 1 (recommended) / keep polishing CONSIDER items / re-audit.
• REVISE / REJECT → take the load-bearing findings (MUST first, then SHOULD) and present them as AskUserQuestion questions, up to 4 per call (batch the rest into a second call after). For each finding:
  • question: the problem in one line, quoting the offending plan text / chapter / seed id.
  • options: 2-4 concrete solutions to THAT problem, the one you recommend first and labelled "(recommended)". Each option is an actual fix (e.g. "Move the apex-tier reveal from ch7 to the climax (ch28)", "Add a Trigger to seed S-04: the dam was already cracked in ch9"), not a vague direction. The author can always pick "Other" to dictate their own fix.

After the author chooses:

• Apply each chosen solution by editing the specific plan/*.md / canon/*.md / setup.md file (you may edit directly once they've picked — the choice IS their consent for that fix).
• Re-run audit_plan.py to confirm the mechanical issues cleared.

Do not re-run the full qualitative pass in this same conversation — go to step 7. The re-audit happens in a FRESH session (the plan changed and this context is now heavy; re-auditing here invites failures).

7. Drop the /clear sentinel and HARD STOP

critique-plan is a token-heavy step (it loads the whole plan + references). Once you have either applied fixes or confirmed PASS, end the session the same way locked chapters do: drop a sentinel so the Stop hook reminds the author to /clear, then STOP unconditionally.

Write the sentinel with a message (the hook uses the file's content as the reminder text):

• If the plan was edited this turn (verdict was REVISE/REJECT):

  printf '%s' '🔧 Plan editado. ESTÁNDAR: ejecuta /clear y reentra con resume-act — el dispatcher volverá a critique-plan para re-auditar en limpio (verdict aún no es PASS).' \
    > output/<series>/book-NN/notes/.clear-pending
  

• If the verdict is PASS with nothing to fix:

  printf '%s' '✅ Plan PASS. ESTÁNDAR: ejecuta /clear y reentra con resume-act (→ escribir capítulo 1 en sesión fresca).' \
    > output/<series>/book-NN/notes/.clear-pending
  

Then print a short closing (verdict, what was edited, path to the critique) and as the last line tell the author:

ESTÁNDAR: ejecuta /clear ahora y reentra con resume-act.

Do not continue into another critique pass or into writing in this conversation. The loop is: critique → fix → /clear → resume-act (→ critique-plan again until verdict PASS, then → write-novel). Because detect_phase only treats critique-plan as done when the verdict is PASS, this loop terminates correctly on its own.

What this skill does NOT do

• Does not modify plan/*.md, canon/*.md, or setup.md without explicit user direction.
• Does not write prose.
• Does not invent plot or characters.
• Does not "fix" the plan — it diagnoses. The user repairs (with you if they ask).