Plan chapter

plan-chapter

You are running the plan-chapter skill. This is the decision gate that runs before any prose is written. Its whole purpose is to put the underdetermined creative choices of the coming chapter in front of the author — with a recommendation — instead of letting the writer improvise them silently. (The skill exists because silent improvisation is how a beam ends up felled by a convenient horse, and how a late reveal gets spent early.)

The output is decisions on disk, not prose. You write notes/decisions-ch<NN>.md; write-chapter reads it as binding.

When to invoke

• The user says "plan chapter N", "briefing for chapter N", "what are the choices for the next chapter?", or is about to write a chapter.
• write-novel chains it automatically before write-chapter (step 2b').
• Re-runnable: if the author wants to revisit choices, run it again; it overwrites the chapter decisions file.

Hard rules

• Decisions, not prose. Never draft chapter text here. You produce a briefing and capture choices.
• Always recommend. Every fork you surface must carry your recommendation as the first option, with a one-line rationale. The author is choosing between informed options, not filling a blank.
• Only surface real forks. A fork is a place the outline/beat sheet does not decide and the writer would otherwise invent: the trigger of an event, who witnesses and how much they understand (exposure), which planted image a payoff resolves, the order of scenes, a tone call. Do not ask about things already fixed by setup, canon, the beat sheet, the seed envelope, or decisions.md. 2-4 forks is the target; a simple chapter may have only 1.
• Honor what is already locked. Read notes/decisions.md (book-level binding law) and the seed envelope first. Never re-ask a settled question; never offer an option that violates a locked decision, canon, or the shadow's reveal timing.
• Protect reveal timing. If a seed's payoff is scheduled for a later chapter, the gate's job is to keep this chapter from spending it. Surface "how loud is the anomaly here?" as a fork with the quiet option recommended.
• No deus ex machina — motivate presence and triggers. Every event's cause and every key character's presence must be intrinsic and seeded, never convenient (no stranger "passing by just when", no fortunate rescue). When a beat needs a character on the page or an event to fire, surface it as a fork and recommend the option where the cause is already on the page and the character has their own reason to be there (ideally tied to the book's engine). If the outline itself smuggles in a convenience, flag and replace it.
• The gate writes; write-chapter obeys. Decisions go to notes/decisions-ch<NN>.md. If a decision is durable (affects more than this chapter — a worldbuilding call, a reveal-timing rule, a name), also append it to notes/decisions.md so it survives a regeneration.

Steps

1. Build / refresh the context bundle

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

--phase plan builds a lighter bundle: it drops the recent chapters in full, the style guide and the craft checklist (the gate decides forks, it writes no prose), so prior chapters arrive as summaries. Everything you need to choose — decisions, canon, neighbor beats, arcs, shadow, the seed envelope and the beat sheet — is still there.

Read notes/_context-chMM.md. Pay special attention to:

• Decisions (the new top block) — book-level locked choices + any prior chapter decisions. These are binding; do not reopen them.
• Beat sheet for chapter M — what is fixed.
• Seed envelope — every seed to plant/echo/pay this chapter, now carrying Trigger, Dose, and Resolution image where present. A payoff with a Trigger field tells you the intrinsic cause is already chosen; a payoff without one is a fork you must surface.
• Shadow slice — what must stay hidden and when each truth is due.
• Arcs — including the exposure ladder (how much each POV's secret is exposed at this point). The current rung is the baseline for any exposure fork.

2. Write the briefing (to the user, in chat)

A short, scannable briefing — not prose. Cover:

• Chapter job — the one structural thing this chapter must accomplish.
• Seeds due — plant / echo / payoff, each with its dose budget and (for payoffs) its trigger and resolution image. Flag any seed whose payoff is ≤2 chapters from its plant: "touch once, quietly; do not re-describe."
• Emotional-spine touch — which planted felt-image this chapter advances or inverts (the "hilado").
• Exposure state — the POV's current rung on the exposure ladder, and whether this chapter is poised to raise it.
• Reveal guards — any later-due truth this chapter could accidentally spend; name it.

3. Identify the forks and ask

List the 2-4 genuine forks. For each, call AskUserQuestion with:

• A clear question naming the stakes.
• 2-4 options. First option = your recommendation, labelled "(Recomendado)", with a one-line rationale in the description. Include the cost/benefit of each alternative honestly.
• Where useful, a preview showing the concrete difference (e.g. an intrinsic-trigger beat vs an external-trigger beat).

Typical fork families (use the ones that apply):

• Trigger of an event — intrinsic & already-seeded cause vs convenient external actor. Recommend intrinsic.
• Exposure / witnesses — who sees the POV's anomaly and how much they understand; how far the ladder moves. Recommend the option that preserves reveal timing unless the outline demands a jump.
• Which image resolves — when a felt-image payoff is due, which planted image it inverts.
• Scene order / emphasis — when the beat sheet lists beats but not their sequence or weight.

Batch independent forks into a single AskUserQuestion call (multiple questions) so the author answers them together.

4. Persist the decisions

Write notes/decisions-ch<NN>.md:

# Decisions — chapter N

> Captured by plan-chapter before writing. write-chapter HONORS these.

- **[fork label]** → chosen: <option>. Rationale: <one line>.
- ...

## Notes for the writer
- Concrete directives that follow from the choices (e.g. "the beam fails from
  the crowd-load + the iron strap finally yielding — both already on the page;
  NO external actor"; "the no-heat anomaly is whispered, not explained — that
  diagnostic belongs to ch24").

If any decision is durable (worldbuilding, reveal-timing rule, a naming call, an exposure-ladder change), also:

• Append it to notes/decisions.md (create the file if absent, using the template in that file's header).
• If it changes the exposure ladder, update plan/arcs.md's ladder for that POV.

Keep canon/seeds edits out of scope here unless the decision is a canon fact the author just settled — in that case make the minimal edit and note it.

4b. Write the continuity contract

Also write notes/continuity-ch<NN>.md — the checkable state sheet the writer must not contradict and the critic checks the prose against. Derive it from the previous chapter's summary carry-forward (in the bundle), the canon character states, and this chapter's shadow slice. List only what is on stage or at risk of being fumbled — 10-20 lines, not a canon dump.

# Continuity — chapter N

> State sheet from plan-chapter. The prose must not contradict a line here.

## Opens at
- Place / time: … (delta from ch N-1 end: …)

## On stage
- <name>: physical state, injuries, what they carry / wear

## POV knowledge
- Knows: …
- Does NOT know yet: … (name the chapter each is due, from shadow/seeds)

## Objects in play
- <object>: where it is, last seen state

This is why continuity errors happen: the canon facts that govern them sit in the low-attention middle of a long bundle. The contract lifts the handful that matter this chapter to a spot right beside the beat sheet.

5. Report and hand off

Print:

• The decisions captured (one line each) and the file path.
• The continuity contract path.
• Any durable decision you propagated to decisions.md / arcs.md.
• The next step: write-chapter N (or, under write-novel, it continues automatically).

What this skill does NOT do

• Does not write or modify chapter prose.
• Does not run critique.
• Does not advance seed statuses (update-canon does that).
• Does not invent plot the author hasn't approved — it offers forks and records the choice.