create-story

create-story: prose that follows on

A model writing a long story keeps the plot and loses the seams. The outline is honoured, the characters are the right ones, and the second paragraph of a scene reads as if it were written by someone who had skimmed the first. Readers feel it as the story losing its train of thought. The research behind this skill (three deep-research reports, September 2026, in docs/deep-research/) traces it to four causes: the whole manuscript in context dilutes the link to the paragraph just written; the model's prior treats a paragraph as a self-contained idea; its prior on relationships resolves tension early; and reasoning run inside the drafting call flattens the prose. Each of the four has a mechanical counter here, and each of the four counters traces to an entry (E1 to E12) in references/evidence.md.

The counters are four files and five scripts rather than a longer prompt. A story bible and a beat sheet fix what each scene is for. A per-scene exit state, in JSON, carries who is where, holding what, feeling what, and what is still open, so the next scene is conditioned on the last one rather than on an impression of the book. A context-pack script assembles the drafter's whole input, including only the last two paragraphs of the scene before, so the window cannot widen by hand. A transition-audit script fails a scene with an orphan paragraph. And critique runs in a fresh context, never in the drafter's.

Whose voice

The skill owns structure and continuity. The voice is composed, and the composition is written down so a critic can quote which card a slip breaks:

• The base is the author. When the author is a named person, invoke that person's content skill first (create-luke-content:create-luke-content for Luke, routed to its book persona) and use its card in references/influences.md as the base; run that skill's lint on every drafted scene at the format the persona names, as well as this skill's checks. When no author is named, the base is a card the user picks or one written for the story during planning.
• Influences are at most three named authors from the catalogue, chosen with the user from the books they say they like. A card carries observable habits (person, tense, sentence-length band, dialogue share, named moves, what to borrow and what to leave), never themes or plots, because an influence is a way of writing and a borrowed subject reads as pastiche.
• Compose, then measure. voice_card.py compose --base "Luke Rhodes" --influence "Andy Weir" --influence "Hugh Howey" prints the ## Voice section: merged bands, at most five rules with their card named, borrow and leave lists, and the sample slot. Paste it into the bible. After each draft, voice_card.py check story/scenes/<id>.md --bible story/bible.md fails a scene outside the bands, and the critic's Voice table quotes a slip against each rule.
• Recall is kept apart from sources. A card's recall: lines come from a model trained on the books (docs/recall-brief.md, run from a plain terminal); compose --use-recall admits them after the sourced moves, each suffixed [recall], so the critic knows which rule rests on a citation and which on memory.
• An author not in the catalogue gets a card before drafting: sourced from interviews, reviews or stylometry that state observable features, with the sources listed on the card, never from memory of the books.