SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsmanuscript-finisher

manuscript-finisher

General & other

View repository

Updated 10-15 20:11

About this skill

Manuscript Finisher

Finishing pass for a manuscript that is content-complete and headed for layout. The point is to spend tokens only where judgment is needed: cheap local tools do the mechanical work; agents do the thinking.

Default style guide: Chicago. Default typography (house): curly quotes/apostrophes (' ' " "), tight em dashes (word—word). Both are configurable in config/.

The four layers (high token-efficiency → high judgment)

Layer 0 — Deterministic report (zero tokens)

Two local engines, no API:

  • LanguageTool (languagetool CLI) — the open-source "Word grammar & spelling." Grammar, agreement, spelling, comma rules, repeated sentence openers.
  • Vale (vale) — prose style + consistency: repeated words, house typography (straight quotes, spaced em dashes), and any custom rules.

Run it:

python3 scripts/run_lint.py <chapter>.md [more.md ...]      # human report
python3 scripts/run_lint.py --json <chapter>.md             # machine report

Output splits every hit into AUTO (mechanical, safe) and REVIEW (judgment). Spelling is filtered against config/allowlist.txt (proper nouns / domain terms — extend it per project; unknown real words flagged once go in here, not into prose changes).

Layer 1 — Mechanical auto-fix (cheap)

  • Typography is fixed deterministically (no agent): 
    python3 scripts/normalize_typography.py <chapter>.md      # straight→curly, spaced em dash→tight, double spaces, space-before-punct
python3 scripts/normalize_typography.py --check <file>.md  # preview counts
  • The agent then applies the remaining AUTO items from Layer 0 (clear typos, obvious agreement slips), and leaves every REVIEW item for Layer 2/the human. Never auto-apply a REVIEW item.

Re-run Layer 0 after fixing; AUTO should drop to ~0, leaving only REVIEW.

Layer 2 — Agentic construal-level passes (parallel subagents)

Spawn one agent per lens over the whole manuscript (not per chapter — these need cross-chapter view). Each returns findings only; it does not edit. Lenses:

  1. Redundancy + internal consistency (macro). Stories/frameworks/quotes repeated across chapters (full retell vs. callback); names/dates/ages/numbers that contradict each other across chapters (e.g. "three years" vs "five years"; a surgeon named two ways; a place spelled two ways). This is the lens the linters cannot do.
  2. Voice + AI-tells (register). Does it sound like the author? Run the project's voice spec (for RLM: rlm-stylometry + de-ai-tell; for non-RLM books the generic AI-tell spec is .claude/skills/clear-writing/references/forbidden-patterns.md). Flag staccato piles, correlative "not just X but Y," manufactured framing, post-quote summary, hedge stacks.
  3. Line-level copyedit adjudication (micro). Take the Layer-0 REVIEW list and rule on each: which LanguageTool/Vale flags are real (apply) vs. false positives or intentional voice (reject, with reason). Enforce Chicago consistency on numbers, hyphenation, serial commas, capitalization.

Orchestrate with the Workflow tool (pipeline by lens) or a parallel() Agent fan-out. Keep each agent read-only; collect structured findings.

Output cap (learned the hard way): the copyedit lens must return a prioritized, capped list (top ~40 APPLY items + grouped REJECT reasons), NOT every lint hit verbatim. A full book can produce 150–200+ REVIEW flags; dumping them all exceeds the 32K subagent output limit and the agent dies silently. For exhaustive coverage, page the REVIEW list per-chapter across several agents rather than one agent over the whole book.

Layer 3 — Synthesis

Merge into one report: applied (what Layer 1 + confirmed Layer 2 fixes changed) and flagged (judgment calls for the author — name, location, the one-line decision). Mirror of the editor's-log pattern. Then the manuscript is ready for the interior build (design-to-book / Paged.js).

Process (typical run)

content-complete chapters
   ↓  normalize_typography.py  (Layer 1 typography)
   ↓  run_lint.py             (Layer 0 report)
   ↓  agent applies AUTO, holds REVIEW   (Layer 1)
   ↓  parallel subagents: redundancy+consistency · voice · copyedit   (Layer 2)
   ↓  synthesis report: applied + flagged   (Layer 3)
   → hand to layout

Config

  • config/.vale.ini + config/styles/Manuscript/*.yml — house typography + repetition rules. Add Chicago/proselint packages with vale sync if network/disk allow (optional; the built-ins + custom rules work offline).
  • config/allowlist.txt — proper nouns/domain terms the spell-checker must not flag. Extend per book.

Notes

  • LanguageTool runs on a markdown-stripped, line-preserving copy so line numbers map back to the source.
  • House style allows em dashes (set tight); the EN_QUOTES/DASH LanguageTool rules are disabled so they don't fight the normalizer.
  • Borrows from: book-quality-loop (judge structure), de-ai-tell + rlm-stylometry (voice/AI-tells; rlm-stylometry is a scoped skill at Creative Intelligence Agency/clients/rlm/.claude/skills/rlm-stylometry/), the cross-chapter redundancy method from RLM book integration.