WriteStory

---
name: writestory-172cd4bea790
evals:
  manual:
    reason: "fiction scaffolding must be judged by the writer for fidelity to their material and usefulness to the intended story"
    falsifier: "the writer finds a proposed layer contradicts their material, flattens a character, or fails to advance the requested story work"
metadata:
  short-description: "Develop fiction from the writer's own material: plot, character, theme, story bible, and prose from short stories to series. Excludes factual summaries."
version: 1.3.3
description: Scaffolds a story the writer already wants to tell, deriving structure, wound, theme, and prose from their own notes across seven narrative layers, short story to series. USE WHEN write story, fiction, novel, chapter, story bible, character arc, plot outline, creative writing, worldbuilding, draft story, help me write my story, develop my novel, layer my story, build out my book, I have notes for a story. NOT FOR narrative summaries of real content, or voice audits and AI-pattern cleanup (Writing).
---

# WriteStory

**Contents:** Ideal State · Intent Routing · The Stance · Tools · Data & Config · Gotchas · Examples

WriteStory is scaffolding for writing fiction. It takes what you already have — a sentence, a folder of notes, a half-drafted chapter — and helps you build the rest of the story one layer at a time, using the systems professional storytellers use on purpose. You bring the creativity; it fills in the pieces you're missing. WriteChapter drafts revisable scenes from your confirmed bible; it never invents the story's spine.

## Ideal State

- Every story element the skill hands back starts from the writer's own material, says whether it was inferred from their seed or offered as a new option, and stands only once the writer ratified or redirected it.
- All seven layers are the destination for every story; the derivation lens the writer's material calls for sets only the order they are reached.
- Each book is a resumable project: its `interview.md` and Story Bible `ISA.md` are read before new work on it and carry what each session confirmed, drafted or changed, so continuity holds across sessions and books.
- Prose comes only from WriteChapter off the confirmed bible, or from Revise on the writer's own draft; either way it was written with the fiction voice model loaded and leaves with its `StoryGate.ts` axes and `StoryJudge.ts` voice line-up shortfall rate attached. The aim is the voice and quality of published fiction, never prose that passes as human.

## Intent Routing

| Intent | Trigger | File |
|----------|---------|------|
| **Interview** | "interview me", "I have notes/an idea for a story", "help me plan a story", "layer my story" | `Intents/Interview.md` |
| **BuildBible** | "build story bible", "create story plan", "map the story" | `Intents/BuildBible.md` |
| **Explore** | "explore ideas", "brainstorm", "I'm stuck", "what if" | `Intents/Explore.md` |
| **WriteChapter** | "write chapter", "write scene", "draft this" | `Intents/WriteChapter.md` |
| **Revise** | "revise", "edit", "improve", "polish" | `Intents/Revise.md` |

The usual path: **Interview** reads what you have and builds the spine with you → **BuildBible** turns it into a Story Bible (a project ISA) → **WriteChapter** drafts off that bible → **Revise** layers and polishes. **Explore** is the idea engine you reach for when a layer is genuinely blank.

| File | Read when |
|------|-----------|
| `StoryLayers.md` | defining or checking any of the seven layers: meaning, character change, plot, mystery, world, relationships, prose (BuildBible) |
| `StorrFramework.md` | building character: sacred flaw, theory of control, the dramatic question, status games (Interview, BuildBible) |
| `PressfieldFramework.md` | concept, hook, theme, villain design and the Foolscap one-page method (Interview, BuildBible) |
| `DerivationLenses.md` | choosing the lens that derives the spine: Storr, Cron, Pressfield/Truby, Egri (Interview) |
| `PhasesAndEvents.md` | mapping three-act structure, beats and mandatory events (BuildBible) |
| `StoryStructures.md` | another framework view: Dramatica, Story Grid, Sanderson, Hero's Journey (BuildBible) |
| `AestheticProfiles.md` | setting genre and style (BuildBible, WriteChapter) |
| `AntiCliche.md` | freshness checks and the mechanisms of machine prose (Explore, WriteChapter, Revise) |
| `RhetoricalFigures.md` | the rare moment that wants a figure, from the Forsyth-based catalogue (WriteChapter, Revise) |
| `Critics.md` | the multi-pass prose review (WriteChapter) |
| `Setup.md` | first-time setup: building the fiction voice model (corpus, bands, exemplars) the drafting gate needs |

## The Stance

WriteStory augments a creator and never substitutes for one. The Ideal State's first bullet (start from the writer's material, label it, let them ratify) is the core of it; these rules carry the rest:

- Every derived element is a proposal the writer ratifies, not a fact the system asserts. Label where each one came from: "inferred from your seed" vs. "a new option I'm offering."
- Offer options, not answers: a few directions to choose among, not one rewrite to approve.
- Suggest at the sentence or beat level. Never take over a chapter.
- Anchor to the writer's own voice: their favorites, their samples, their cadence.
- The writer's word always wins. If they redirect, the redirect is the truth.
- Only WriteChapter emits new prose, and only off the writer's confirmed bible, framed as a revisable draft they own; Revise works only on the writer's own draft.

## Tools

| Tool | Purpose |
|------|---------|
| `Tools/BuildFictionCorpus.ts` | Splits Project Gutenberg plain-text story collections into per-story `.md` files plus `Corpus.md` manifest. |
| `Tools/BuildVoiceProfile.ts` | Measures every corpus piece and records p10/p50/p90 bands per metric in `VoiceProfile.json`; `--compare <file>` checks a draft against them. |
| `Tools/StoryGate.ts` | Measures draft prose against a human-fiction corpus (dialogue, beat rate, universals) and flags machine-prose drift. |
| `Tools/StoryJudge.ts` | Blind voice line-up against corpus paragraphs (the verdict: does the draft hold up in voice and craft) plus a six-mechanism diagnosis against `Exemplars.md` (advisory); `--detect` adds a which-was-model rate as a diagnostic only. |
| `Tools/VoiceMetrics.ts` | The deterministic metric vector the other tools share; `bun Tools/VoiceMetrics.ts <file> --pretty` prints it for one text. |

## Data & Config

First-time setup (accounts, keys, files to fill in): [Setup.md](Setup.md).

The fiction voice model and every book project are USER data the skill reads at run time; the skill body holds none of it.

| What | Where |
|---|---|
| Corpus config (Gutenberg ids, authors, anchor) | `<configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/WriteStory/Corpus/Corpus.json` |
| Per-story corpus files + manifest | `…/WriteStory/Corpus/<slug>.md`, `…/WriteStory/Corpus/Corpus.md` |
| Bands | `…/WriteStory/Corpus/VoiceProfile.json` (from `Tools/BuildVoiceProfile.ts`) and `…/WriteStory/Corpus/StoryProfile.json` (fiction axes, from `StoryGate.ts --build`) |
| Exemplars | `…/WriteStory/Corpus/Exemplars.md` — verbatim passages, loaded before drafting and by the judge |
| History (optional) | `…/WriteStory/History.md` — the install's own incident specifics behind generic lessons here, read on demand |

Build order on a fresh install:

```bash
bun ~/.claude/skills/WriteStory/Tools/BuildFictionCorpus.ts --config <Corpus.json> --fetch
bun ~/.claude/skills/WriteStory/Tools/BuildVoiceProfile.ts --manifest <Corpus.md> --out <VoiceProfile.json>
bun ~/.claude/skills/WriteStory/Tools/StoryGate.ts --build --manifest <Corpus.md> --voice <VoiceProfile.json> --out <StoryProfile.json>
```

Then, per draft: `StoryGate.ts <draft>` (deterministic bands plus the narrator's aside) and `StoryJudge.ts <draft>` (the verdict is a blind voice line-up of the draft's narrative paragraphs against corpus paragraphs, a judge asked which carries the stronger voice and craft, shortfall rate at or under 0.6 to pass; the six-mechanism diagnosis it prints is advisory, and the `--detect` model-detection rate is a diagnostic that never decides a pass). Every tool exits 2 with the build instruction when the model is absent.

Everything above lives under `<configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/WriteStory/` (`configRoot` is `$CLAUDE_CONFIG_DIR` when set, else `~/.claude`). When that directory holds `PREFERENCES.md`, its default genre, aesthetic and voice apply.

Each book is a project under `LIFEOS/USER/CUSTOMIZATIONS/SKILLS/WriteStory/projects/<book-slug>/`:

- `interview.md` — the saved interview results (the spine, the layers, what's confirmed)
- `ISA.md` — the Story Bible as a project ISA: the living plan that tracks what's done and what's left, growing across sessions

## Gotchas

- **The words that fail are ordinary.** A story that passes every banned-word table can still read as machine prose through narrator's asides ('the way everyone reads agreements', 'as anyone would') and rhythm tells; `AntiCliche.md` category 7 is the pass that catches them, and it is a read, never a grep.

## Examples

**Example 1: A writer who already has a lot**
```
User: "I've got a folder of notes for a novel — a few character sketches,
       a rough world, and two draft scenes. Help me build it out."
→ Interview intent
→ interview.md saved: a spine derived from their own notes, every element labeled and ratified, all seven layers marked confirmed / partial / open, the next step left to them
```

**Example 2: Building the full story plan**
```
User: "Build the story bible for my novel"
→ BuildBible intent
→ The project ISA maps all seven layers start to finish, its criteria track done versus remaining, and the gaps still needing a decision are named
```

**Example 3: Writing actual prose**
```
User: "Write chapter 3 based on the story bible"
→ WriteChapter intent
→ A draft of chapter 3's beats the writer owns and revises, reported with its gate axes and a voice line-up shortfall rate at or under 0.6
```