Storygen

Storygen: Agentic Story Generation

We've taught AI to code, plan, and reason. Making good stories is becoming easier; making great stories is still hard, for three reasons this skill is built to counter:

• Data. Exceptional storytelling data is scarce — and the development drafts and revision histories matter as much as the finished text. The process matters as much as the output, so this pipeline captures the iteration: plans, notes, and revision logs are artifacts, not scratch.
• Subjectivity. There is no single "right" answer. Good storytelling is judged on four axes — coherence, emotion, persuasion, resonance — so testing uses a panel of specialist reviewers (simulated editorial feedback), not a single score.
• Optimization. It's easy to optimize the wrong thing. The target is deeper narrative craft — reader immersion, empathy, transformation — never style simulation. Convergence rules below exist to stop polish-loops that improve prose while the story stays broken.

The setup mirrors a coding agent: coding runs Plan → Write → Test → Debug → Refactor; storytelling runs Plan → Write → Test → Revise. And like software, a story is a living system of contexts — it breaks when state isn't updated correctly — so all phases read and write a story bible.

Two North Stars, from Pixar's Braintrust:

1. Candor without authority. Feedback is advisory, never prescriptive. Reviewers diagnose problems; the writer owns the fix. This preserves creative agency.
2. Make the story better, not be right. Optimize whole-story quality, never an individual component's score.

Pipeline

Copy this checklist into a todo list and track progress:

Storygen Progress:
- [ ] Phase 0: Intake — brief, audience, format, constraints
- [ ] Phase 1: Plan — concepts, arc, spine, story bible, beat outline
- [ ] Phase 1.5: Storyline checkpoint — Braintrust reviews the outline
- [ ] Phase 2: Write — full draft, scene by scene
- [ ] Phase 3: Test — Braintrust reviews the complete draft (parallel)
- [ ] Phase 4: Revise — triage notes, rewrite, update bible
- [ ] Phase 5: Converge — repeat Test/Revise until pass, then deliver

Phase 0: Intake

Establish before planning (infer sensible defaults from the request rather than interrogating the user; ask only if the request is genuinely ambiguous):

• Premise — the seed idea, in the user's words. It can be tiny ("I have a crush on this guy at school"); the pipeline expands it.
• Audience & purpose — a 6-year-old's bedtime tall tale, a literary short story, a teaching story (e.g. learning to apologize), a screenplay scene.
• Format & length — prose, script, storybook/picture-book pages, comic or storyboard panels; word or page count.
• Constraints — real people or details to include, canon rules if remixing existing universes, tone boundaries, content limits.

Phase 1: Plan

Never draft prose before the plan exists. Craft standards are in craft.md — read it before planning. Produce, in order:

1. Concepts. If the premise is open-ended, generate 3–5 candidate concepts first — same seed, different genre lenses and framings — then pick one (or present them if the user wants to choose). A one-line daydream can expand into a dozen distinct story concepts; committing to the first idea is how stories end up generic.
2. Logline — one sentence: protagonist + desire + obstacle + stakes.
3. Arc choice — pick the arc shape that fits the story from the seven in craft.md (Classic/Freytag, Hero's Journey, Three-Act, Fichtean, Circular, Braided, Kishōtenketsu). Default: Classic Arc. State why.
4. Story spine — fill in every slot:

   Once upon a time there was… Every day… One day… Because of that… Because of that… Until finally… And ever since then… Each "Because of that" must be causal, not sequential. If you can swap two beats without breaking anything, the spine is broken.

5. Story bible — create the living state document from the template in story-bible.md: world seed (environment type + defining twist + inhabitants + power/resource + struggle/conflict), character sheets (desire, need, flaw, strengths, weaknesses, core quality, arc), world rules, timeline, scene ledger, open threads.
6. Beat outline — one line per scene: who wants what, what turns, what changes state. For visual formats (storybook, comic, screenplay), take it down to a shot/panel list per scene. Every scene must earn its place by changing something.

Non-negotiables: the protagonist is driven by desire and conflict; the audience admires the effort, not the success; characters are forced to grow through uncomfortable situations.

Phase 1.5: Storyline checkpoint

Pixar's Braintrust reviews storyboards before animation — structural bugs are cheapest before prose exists. Run a light Braintrust pass on the outline: give the reviewers the spine, bible, and beat outline. Structure, character, tension, and worldbuilding lenses apply fully; emotion and voice review what's visible at this stage (planned emotional arc; planned voice notes). Fix BLOCKING findings in the plan before writing a word of prose.

Phase 2: Write

Draft the full story from the beat outline. While writing:

• Work scene by scene; each scene has a turn (a value that flips: hope→fear, ignorance→knowledge, safety→danger).
• Keep voice distinctive and dialogue purposeful — every line either reveals character or advances the story, ideally both.
• Honesty in the unbelievable: however fantastical the premise, emotions and choices must be true. A dragon may be impossible; the fear of disappointing your father is not.
• After drafting, update the story bible: what each character now knows, wants, and has lost; which threads opened or closed. Stale state produces continuity bugs.

Phase 3: Test — the Braintrust

Run the multi-agent review. Full protocol in braintrust.md.

If subagents / a Task tool are available, launch these six reviewers in parallel, in a single batch, each readonly, each given the complete draft plus the story bible:

If subagents are not available, run the same six reviews sequentially in-context using the condensed rubrics in braintrust.md, adopting one persona at a time. Do not skip lenses.

Phase 4: Revise

1. Collect all notes. Discard none unread.
2. Triage against the whole story, not the note's severity alone: a note is only actioned if the fix makes the story better. When notes conflict (e.g. tension wants a darker midpoint, emotion wants warmth), the story spine and the protagonist's arc are the tiebreakers.
3. Rewrite the affected scenes — fix root causes (usually in structure or character motivation), not symptoms (usually in prose).
4. Update the story bible — including its revision log — to reflect every change. Then re-verify the spine still holds end to end.

Phase 5: Converge

Re-run the Braintrust on the revised draft. Ship when:

• No reviewer reports a blocking note, and
• Two consecutive rounds produce only polish-level notes, or
• Three full rounds have run. Style simulation is not narrative craft: a fourth pass of prose-polish cannot fix anything that matters. If round three still has blocking notes, the bug is in the plan — return to Phase 1 and rebuild the spine.

Deliver the final story plus its iteration history: the logline, the chosen arc and why, the spine, and what changed across revision rounds. The process matters as much as the output — the development drafts and revision history are part of the deliverable, not scratch to discard.

Additional resources

• craft.md — what makes stories great: spine, arcs, character, layers, world seeds, format playbooks (tall tales, teaching stories, universe remixes, storyboards)
• braintrust.md — review protocol, rubrics, note format
• story-bible.md — living state template and update rules