Close act

close-act

You are running the close-act skill. The act is finished. Your job is to close the conversation cleanly: compress the act's summaries, stabilize voice observations, write the handoff document, and tell the author it's safe (and recommended) to /clear before the next act.

After this skill runs, the next session can rebuild everything from disk via resume-act.

When to invoke

• The author says "close act N" / "wrap up the session" / "cerramos el acto".
• update-canon for the last chapter of an act has finished. The update-canon skill itself prints Strongly recommended: run close-act then /clear when this condition is met.

Hard rules

• Seeds and shadow are NEVER compressed. They are the source of truth for foreshadowing and the writer's hidden timeline. Closing the act does not touch them.
• Voice consolidation is stabilization, not invention. You can only consolidate what voice.md actually says. Do not promote a one-time observation to a stable rule unless it appeared multiple times during the act.
• The handoff must be self-contained. A reader (future you in a fresh session) should be able to pick up the next chapter armed with grimoire + plan + canon + handoff alone. No reliance on conversation memory.
• End with the explicit signal. The last line of your final response must be: ✓ Act N closed. STANDARD: run /clear now, then resume-act for act N+1.

Steps

1. Run the helper

python3 .claude/skills/close-act/scripts/close_act.py \
    --series-slug <slug> --book-number <N> --act <A>

This:

• Runs prepare_act.py (bundles the act's chapter summaries into notes/_act-AA-bundle.md + writes a skeleton at summaries/act-AA.md for you to fill).
• Resets notes/session-handoff.md to a fresh skeleton dated today, with the four TODO sections you'll fill below.

2. Fill the act summary

Read the bundle at notes/_act-AA-bundle.md. Fill summaries/act-AA.md to about 1500 words. Lose texture; keep facts.

Act-summary structure: one or two paragraphs of what happened (plot, in order), then a short list of the durable consequences the later chapters depend on — promotions to canon, relationship shifts, locations established, seed plants/payoffs that occurred. Drop scene texture and dialogue; keep anything a future chapter must not contradict.

2b. (Re)build the book summary synthesis

update-canon no longer rewrites summaries/book-summary.md every chapter (it only touches the ## What just happened line). The full synthesis is built here, once per act — this is what makes the book summary effectively incremental by act instead of per-chapter.

Open summaries/book-summary.md and (re)build it from the act summaries (summaries/act-*.md) plus the current canon state. Target ~2000 words, ruthless. Structure:

# <Book title> — running summary

## State of the world right now
> 4-6 bullets: political, magical, geographic state at the most recent act.

## State of principals right now
> One line per principal: location, emotional state, goal, cost borne.

## Threads in motion
> 3-8 lines: open subplots, unpaid seeds (by id, not content), debts.

## Reader's accumulated knowledge
> 4-6 lines: what the reader now believes (may differ from shadow truth).

## What just happened
> 2-3 sentences on the most recent chapter (carry over / refresh whatever
> update-canon left here).

Source of truth is the act summaries, not the conversation. If it grows past ~2500 words, trim the oldest threads. This file is what becomes the previous-book context for the next book in the trilogy.

3. Consolidate voice.md (stabilization)

Open notes/voice.md. Walk through the entries added during this act (date-stamps within the act window). For each POV section:

• Group observations that say similar things.
• Identify the rules that held across ≥ 2 chapters of the act. Those are stable.
• Identify rules that were stated once and didn't reappear. Those are provisional — leave them in the per-entry log but do not promote.

At the end of each ## POV: <name> section, add or update a sub-section:

### Stable rules (after act {act})
- (consolidated rule 1)
- (consolidated rule 2)
- ...

If a previous act already wrote a Stable rules block, update it instead of duplicating. Older rules stay unless contradicted by new ones — and contradictions go to open-questions.md for resolution.

Then consolidate the "Recurring patterns to watch / avoid" list the same way — and keep it short, because the critique's 8d sweep must hunt every entry by eye, and an LLM's exhaustive-checklist reliability collapses past ~7 items:

• A pattern that is countable by regex (a phrase form, a word, an opener shape) graduates: add it to notes/prose-lint.toml (as an [[extra_patterns]] entry, or [[reserved]] if it is a reserved word) so it is counted deterministically, and replace its voice.md entry with a single line: "→ graduated to prose-lint". This is the whole point of the prose auditor — move counting off the critic's eye.
• Judgment-only patterns (ones no regex can catch — an aphoristic close, an erudite aside that leaks a reveal) stay in voice.md, capped at 7. If the list would exceed 7, merge overlapping entries or demote the weakest; a 12-item sweep silently checks none of them well.

3b. Fold style-rules.md into style.md

notes/style-rules.md is a capture buffer — prose rules the author stated in chat that aren't yet in the book's style guide. Fold them now so there is one source of truth:

• For each entry under ## Unfolded, edit style.md to express the rule in the right section (prose register, reveal timing, dialogue…), updating an existing bullet rather than duplicating.
• Skip process rules (word-count floor, expand-chapter discipline) — those belong to the write-chapter skill, not style.md. If the buffer contains one, just drop it (the skill already owns it).
• After folding, reset style-rules.md to its empty template (the ## Unfolded list back to - (no rules declared yet)).

3c. Consolidate decisions.md (prune what canon now enforces)

notes/decisions.md accumulates book-level binding choices and is loaded into every chapter bundle, so it must not grow without bound. At each act close, walk it and prune entries that are now enforced elsewhere:

• A decision whose substance is already promoted to canon/* or setup.md (a name, a worldbuilding fact, a magic cost) is redundant here — canon already enforces it on every chapter. Remove it from decisions.md.
• A decision that is still load-bearing but not yet in any canon file (a reveal-timing rule, a craft constraint, an exposure-ladder choice) stays.
• When in doubt, keep it — only prune what you can point to verbatim in a canon file. Never drop a decision you cannot confirm is enforced elsewhere.

Also delete notes/decisions-ch<NN>.md files for chapters in this act only if their durable items were already appended to decisions.md (per plan-chapter's rule); otherwise leave them. This keeps the per-chapter gate files from piling up while preserving anything not yet consolidated.

3d. Curate voice exemplars

notes/voice-exemplars.md holds the author-blessed prose the writer imitates — the only fixed in-voice sample in the write bundle (without it, the writer's only in-voice text is the previous chapter's tail, so a tic there photocopies forward). Refresh it from the act's best prose:

• Scan this act's chapters for 2-3 candidate passages of 150-300 words that best embody the book's voice — the natural pool is prose the chapter critiques praised in their ## What works sections. Prefer passages that show the hard things (sensory texture that carries plot, emotion underplayed, a seed planted obliquely) over merely pretty ones.
• Present the candidates to the author with one AskUserQuestion (per candidate: keep / replace an existing exemplar / skip). Only the author's confirmed passages go in.
• Write the confirmed passages to notes/voice-exemplars.md, capped at 5 total — this is REPLACE, not accumulate: if adding would exceed 5, drop the weakest existing one. Quote verbatim; label each with what it demonstrates and its source chapter.

4. Write the session handoff

Open notes/session-handoff.md (just reset by step 1). Fill the four TODO sections, in order:

4a. Voz consolidada

The synthesis from step 3. 3-6 stable rules per POV, expressed as prescriptive sentences ("Bruno habla con frases cortas. Pausa antes de contestar. Verbos artesanales sobre verbos de pensamiento."). This becomes the fast read for the next session — the voice.md file is the long form, this is the executive summary.

4b. Hilos abiertos para el acto siguiente

From notes/open-questions.md (## Pendientes section), copy the items that the next session must address. Prioritize: anything blocking the next chapter goes first.

4c. Decisiones de esta sesión

List the decisions taken in chat during this act that affect future chapters. Be tight: one line each, with origin. Examples:

• "(ch 4) Cambiamos el POV original de Bruno a Vela para que el plant de la Lectura no quede telegrafiado."
• "(ch 6) Mentor reformista renombrado: 'Doral'."

Only include decisions not already captured elsewhere (plot decisions should be in setup.md; name decisions should be in canon).

4d. Notas para resume-act

What does the next session need to read FIRST, before opening the context bundle? If there's a critical gotcha for chapter N+1 (e.g., "chapter N+1's POV is Yedra, not Bruno — easy to forget"), put it here. If nothing critical, leave a one-liner: "No special notes; standard resume-act bundle suffices."

4e. Naive-reader calibration (measure telegraphing)

The critique can't tell whether a reveal is telegraphed — it knows the whole shadow, so it can't feel what a first-time reader picks up (the curse of knowledge). A naive-reader subagent can: it reads only the prose and reports what it actually suspects. Ask the author with one AskUserQuestion: "Dispatch the naive reader for acts 1-N?" On yes:

1. Dispatch the naive-reader agent (Agent tool). In the prompt, enumerate the exact chapter file paths so far — output/<series>/book-NN/chapters/01.md … NN.md — as an explicit list, not a glob. The enumerated-paths-only rule is what keeps the reader plan-blind; never hand it a directory or a glob.
2. Write its returned report verbatim to notes/reader-report-act<NN>.md (committed — it is a record).
3. Then do the comparison the subagent structurally cannot (it never saw the plan). For each high-confidence suspicion or prediction, hold it against plan/shadow.md: find the truth it targets and its Reveal cap / current Status. Append a ## Telegraphing assessment (author-side) section to the report — flag any truth the reader reached with high confidence while the shadow reserves it for far later (e.g. "predicted high-confidence at act 1 but naturaleza-de-bruno is capped sensed until ch 22 → telegraphed", quoting the reader's passages). Copy anything actionable into notes/open-questions.md for the next session.

This step is optional per act (skippable when the author declines) and never edits the plan or prose — it only measures and reports.

5. Verify nothing critical was missed

Before signaling done, check:

• summaries/act-NN.md filled (~1500 words).
• summaries/book-summary.md rebuilt (~2000 words) from the act summaries (step 2b).
• notes/voice.md has a Stable rules (after act N) block per POV that appeared.
• notes/session-handoff.md has all four TODO sections filled.
• notes/voice-exemplars.md refreshed (≤5 author-confirmed passages).
• plan/seeds.md and plan/shadow.md untouched (open them to verify if unsure).
• notes/open-questions.md reviewed — pending items either copied to handoff or explicitly de-prioritized.

6. Report and signal

Print:

✓ Act N closed.

- summaries/act-NN.md: <word count>
- summaries/book-summary.md: rebuilt (<word count>)
- voice.md: stable rules consolidated for POVs: <list>
- voice-exemplars.md: <N> author-blessed passages
- session-handoff.md: <date>, with N pendientes for next session

Next session bootstrap: run `resume-act` (act N+1, chapter <next>).

STANDARD: run /clear now, then resume-act for act N+1.

The last line is the explicit signal. Do not add anything after it.

What this skill does NOT do

• Does NOT touch plan/seeds.md, plan/shadow.md, plan/outline.md.
• Does NOT delete chapter summaries (summaries/ch-NN.md) or chapter prose. They stay on disk for search and inspection.
• Does NOT modify canon (canon is locked by update-canon per chapter).
• Does NOT run /clear — that's a user command. This skill signals that /clear is safe and recommended; the author executes it.