Story Bible Continuity

Long manuscripts fail in a specific, boring way: the writer (or the model) remembers the story as a vibe, not as facts. A character is 18 in chapter 1 and 23 three years later in chapter 12. A name drifts one character and stays drifted. A phone appears in 1998. A planted key is never paid off, and the author discovers it at chapter 60 while the reader noticed at chapter 20.

Re-reading the draft does not catch these, because re-reading is the same memory that produced them. The fix is a story bible — the story's facts as data — plus a checker that compares prose against it mechanically. Facts are cheap to check. Vibes are not.

This skill is the checker and the adjudication discipline around it. It does not draft, outline, or polish. It answers one question per hit: is the manuscript wrong, or is the bible stale?

When to Use

Use this skill when:

• a manuscript or serial passes ~30k characters / ~10 chapters and you need a continuity pass before publishing or before the next arc;
• the writer reports "a name feels off", "I lost track of someone's age", or "I forgot whether I already paid off the key";
• you are about to continue a draft that was paused for weeks and the writer's memory of the earlier chapters is no longer reliable;
• a model-written draft has to be checked against facts the writer established earlier, not against what the model believes the story is.

Do not use it to judge quality, pacing, voice, or plot logic — see Limitations.

The bible

One JSON file beside the manuscript. Per-character birth years, era bounds and forbidden terms, allowed POV characters, and a foreshadowing ledger with planted/payoff chapters. Full schema, field-by-field, with a worked example: references/story-bible-schema.md.

{
  "title": "衣冠",
  "era": {"setting_period": "1998-2005", "forbidden_terms": ["智能手机", "微信"]},
  "pov": {"mode": "third-limited", "allowed_pov": ["修逸"]},
  "characters": [
    {"name": "修逸", "aliases": ["小修"], "birth_year": 1979,
     "watch_spellings": ["修毅"]}
  ],
  "foreshadowing": [
    {"id": "F1", "summary": "父亲留下的钥匙", "planted_chapter": 3,
     "payoff_chapter": null, "payoff_keyword": "钥匙", "deadline_chapter": 40}
  ]
}

If no bible exists, build one first: python3 sbc.py init bible.json --title ... --period ..., then fill characters and the foreshadowing ledger from the chapters that already exist, not from what the story is supposed to be. A bible written from intent hides exactly the drift you are looking for.

The check

The checker is shipped as source, not as a committed script: copy the Python block in references/sbc-checker.md to sbc.py beside the bible (stdlib only, Python 3.8+), then:

python3 sbc.py check bible.json manuscript/          # human report
python3 sbc.py check bible.json manuscript/ --json   # machine report
python3 sbc.py check bible.json manuscript/ --fail-on warn

Exit 0 = clean or warnings only, 1 = errors, 2 = usage. Chapter files are recognised as ch01.md, 003_draft.txt, 第3章.md — anything with a number, .md/.txt, sorted by that number.

The adjudication

Every hit is one of three things, and the skill decides which — never silently picks one:

1. Manuscript bug → fix the prose. The bible is the record of what earlier chapters already committed to, so the newer chapter yields.
2. Stale bible → the newer chapter changed a fact on purpose. Update the bible, and say so explicitly: which field changed, from what, to what, and which earlier chapters now conflict.
3. Real retcon → the story genuinely overrides an earlier fact. That is allowed, but it is a decision, not an accident: list every earlier passage that now needs rewriting before touching either file.

Report format per hit: file:line, the check name, the two conflicting facts, and the verdict. Batch them by verdict, not by file, so the user sees "3 prose fixes, 1 bible update, 1 retcon to approve" instead of a wall of lines.

Limitations

Say this out loud rather than implying full coverage. It does not judge plot logic, motivation, pacing, voice, or whether a name is deliberately reused. name-drift only fires on declared spellings and on speaker position, so a mistyped name mid-sentence passes. pov-leak is heuristic. Continuity that needs semantic judgement — "would this character know that by now?" — is a reading task, and the honest answer is to read the two chapters side by side, not to claim the script covered it.

Never edit a manuscript to make a check pass. A checker that can be satisfied by rewriting evidence is worthless.

Verification

Before reporting a manuscript as clean, prove the checker runs: make a throwaway two-chapter fixture (ch01.md, ch02.md) with one deliberate name-drift and one missing payoff, run python3 sbc.py check, and confirm it reports exactly those two issues and exits 1 (deterministic, ~1s, no model).

Then run python3 sbc.py check on the real manuscript and paste the summary line (N error(s), M warning(s)) with the verdicts you assigned. A clean run on an empty folder, or a run whose only output is "0 errors" on a bible with no characters, proves nothing — check that the bible actually has entries and the chapter count matches the folder.