Draft Scene

Skill: draft-scene

What it does. Generates scene prose from a validated outline contract, carrying the full weight of the project's knowledge: persona, references, style guide, world rules, character voices, preserve list, recurrence map, and the 14-category framework as negative constraints. Context-expensive by design — the point is prose born inside the voice, not prose cleaned afterwards.

Triggers

• "draft scene 2 of chapter X.Y"
• "/draft-scene X.Y scene-2"
• "write the scene where {...}" (only if an outline covers it)
• Equivalent phrasing in the project's output language

Input. A scene from a validated {chapter}_outline.md, or an ad-hoc brief containing at minimum: purpose, conflict, turn, who is on stage, protection constraints.

Output. {chapter}_draft.md (or appended scene block), explicitly marked as a draft for curation. Never written into the manuscript file itself.

Report the draft's length against the scene's budget. If the chapter file carries ## n - name - budget headers, scripts/scene-budget gives the surrounding context; if it does not, use the outline. Never add those headers to the author's chapter, and never note that they are missing — see manuscript/README.md → Scene headers while drafting.

Write the prose in the project's output language.

Core principle

The draft is a proposal. The measure of success is how much survives the author's curation.

This skill does not produce finished prose. It produces the best possible starting material — prose that already respects the voice, the world, and the constraints, so the author's revision energy goes into making it theirs instead of removing the machine.

Expected trajectory: early drafts survive curation at maybe 40–60%. As persona.md, preserve-list.md, and revision-log.md accumulate, survival should rise. If it does not, the persona is not yet sharp enough — say so and propose a define-persona update session.

Hard rules (before anything else)

1. No contract, no draft. If there is no validated outline covering the scene and the author has not provided a brief with purpose/conflict/turn/protection, stop and build one first. Drafting without a contract produces plot, and plot decisions belong to the author.

2. Protection constraints are absolute. The outline's "must NOT happen" list and worldbuilding's deliberately-unexplained list are inviolable. If drafting seems to require touching one, stop and surface the conflict.

3. Open questions are hard stops. Where the outline says "the author hasn't decided", the draft ends or routes around. Never invent the undecided.

4. Never touch the manuscript file. Output goes to a _draft file. The author moves text into the manuscript; the skill does not.

Execution protocol

Step 1 — Full context load

Read, in this order (yes, all of it — this is the expensive part and the entire point):

1. .project/config/project.yaml — output language
2. The scene contract — the outline entry or ad-hoc brief
3. The previous scene/chapter — continuity of action, tone, and story time
4. .project/config/persona.md — the voice. Re-read the model passages (section 8) immediately before drafting; they are the tonal calibration. Note the anti-models (section 9) as what to avoid. Note the per-1k ceilings in "signatures to watch" (section 4)
5. .project/config/references.md — borrowing instructions and their exclusion lines
6. .project/config/style-guide.md — hard formatting and vocabulary rules, plus the project's ceiling overrides
7. .project/knowledge/worldbuilding.md — active rules, costs, knowledge distribution, deliberately-unexplained list
8. Character sheets for everyone on stage — Voice sections above all; sample lines are calibration
9. .project/reports/preserve-list.md — thesis phrases (may echo intentionally only if the contract calls for it)
10. .project/reports/recurrences.md — phrases already used; never generate them again
11. .project/templates/framework.md — the 14 categories as negative constraints

Step 2 — Compile the constraint sheet

Before writing a word, produce (internally) the scene's constraint sheet:

• Voice targets: register, rhythm, the 2–3 persona devices active in this kind of scene
• Density budget: convert the ceilings into an absolute allowance for this scene's target length. Ceilings are per-1k (persona §4 and style-guide.md overrides, otherwise the framework defaults); a 900-word scene against a 2,0/1k ceiling gets 1,8 → 1 occurrence. Round down: a scene is not entitled to its full share of every category at once. Write the allowance out per category before drafting, or the self-audit in Step 4 has nothing to check against
• Character voice keys: one line per character on stage — their register, their never-says
• World guards: which rules are active in this scene; who on stage knows what
• Protection list: what this scene must not reveal or resolve
• Recurrence blocklist: striking phrases from the map that must not reappear
• Length target: from the contract's length signal

Step 3 — Draft

Write the scene against the constraint sheet.

Guidance that matters most in practice:

• Purpose discipline. Every beat serves the scene's stated purpose. Beautiful digressions that serve nothing get cut before the author ever sees them.
• Emotion through the body. Category 14 is the most common generation failure. Gesture, breath, timing — not labels.
• Dialogue asymmetry. Category 9's inverted problem too: vary reply lengths, one gesture between lines, not two.
• Endings are not punchlines by default. Check what the recent chapters' scenes did; vary.
• When in doubt, under-write. The author can expand a lean draft; deflating an overwritten one costs more.

Step 4 — Self-audit (before the author sees anything)

Run the draft through the 14 categories as if it were a hostile analyze-chapter pass:

• Count the drafted words, then count each category against the Step 2 allowance
• Anything over budget → revise now, and re-count after revising
• Check every striking phrase against the recurrence blocklist
• Check protection constraints one final time
• Check each character's lines against their never-says

The author should never receive a draft that would fail the project's own analysis. If a tic survives self-audit because it genuinely serves the scene, keep it — but flag it in the delivery notes.

Step 5 — Deliver with honest notes

Present the draft with a short delivery note:

## Draft notes — {scene}

**Self-audit:** {N,N}/1k over {N} words — {clean | N flags kept deliberately, listed below}
**Choices you should review:**
- {a voice call, a pacing call, an interpretation of the contract — anything the skill decided that the author might decide differently}
**Where I stopped:** {open questions respected, boundaries hit}
**Length:** {target vs. delivered}

The "choices you should review" list is mandatory. A draft with zero flagged choices means the skill is hiding its uncertainty, not that it had none.

Step 6 — Iterate on feedback

When the author reacts ("she wouldn't say that", "too fast here", "this whole beat is wrong"):

• Revise only what the feedback touches. Do not silently rewrite untouched passages.
• Every voice correction is data: propose recording it — character sheet Writing notes for "she wouldn't say that", persona for style corrections.
• If the author rewrites a passage themselves, their version wins and becomes reference material for the next scene's calibration.

What to avoid

Drafting to impress. The failure mode of every generation pass: reaching for the striking phrase. The persona's model passages define what "good" means here — not general literary ambition. Restraint survives curation; flourish gets cut.

Resolving what the contract left open. The most expensive failure. When the draft needs an answer the author hasn't given, the draft stops.

Recycling the book's own hits. A phrase that worked in chapter 3 is not available in chapter 9. That is what the recurrence blocklist is for.

Uniform pacing. Scene contracts carry length signals. A "brief" scene drafted at full weight breaks the chapter's shape.

Silent confidence. Every interpretive choice goes in the delivery notes. The author curates; the skill discloses.

Relationship to other skills