convez.ai
Skill
Quentinbest

McKee-stroy-workflow

Quentinbest/McKee-stroy-workflow

Ver no GitHub
InicialPluginScore · de 0 a 100

McKee-native story generation platform: 21 Skills + 22 Agents for reliably producing great stories with Claude Code

Instalação

$npx convez-ai Quentinbest/McKee-stroy-workflow

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 2 meses
Versão
—
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

McKee Story Workflow

A McKee-native, hybrid Skill + Agent architecture for consistently and reliably producing great stories using Claude Code.

Built on Robert McKee's Story methodology. See story-plan.md for the full implementation design.

Note: the repository slug is misspelled McKee-stroy-workflow (stroy → story). The remote URL still uses the typo; consider renaming the GitHub repo (and updating clones' remotes) when convenient. The skills themselves are unaffected — they install by their own names into ~/.claude/skills/.

Cross-platform note

The skills are authored for Claude Code (where critics run as parallel isolated agents). They also run on hosts without an agent-spawning tool — OpenCode, Pi, etc. — by degrading gracefully: the critic suite falls to native critic tools (where the host provides them, e.g. Pi's cliche_hunt/subtext_check) or to in-context sequential passes with a fresh-eyes reset between dimensions. See /story-audit Step 0 (capability ladder) and its Cross-Platform Critic Map. On non-Claude-Code hosts the skill body is typically pasted/loaded as a prompt; the capability ladder is what makes that portable.

story-scene now also uses an internal Beat Gate before full-scene critics: deterministic mechanical fixes when provably safe, blind Beat critique, one consolidated writer decision, resumable scene ledgers, and rolling reader/pacing checks every 2-3 committed scenes. Hosts without Node.js or command execution fall back to detect-only and never self-authorize AUTO.


What This Is

A complete AI-agent workflow for story generation, organized into four layers:

skills/mck-*          McKee methodology skills (the how-tos)
skills/story-*        Workflow skills (the lifecycle entry points)
agents/               Generator + critic agents (bounded, isolated work)
templates/            Project scaffolding (lifecycle.json, state.json)

Skills

Methodology Skills (/mck-*)

Drop into ~/.claude/skills/ to make available globally in Claude Code.

V1 — Core

SkillPurpose
mck-subtext-5layerAuthor dialogue in 5 layers (Wound → Want → Fear → Tactic → Text). The most important single skill.
mck-controlling-ideaForge the story's "value + cause" sentence
mck-beat-to-proseTranslate a beat sheet into polished prose
mck-crisis-dilemmaSharpen a hard choice into a true dilemma
mck-arc-walkMap a character's arc across the spine
mck-gap-findFind the Gap (expectation vs. result) in any beat or scene

V2 — Long-Form & Polish

SkillPurpose
mck-image-threadInventory motifs, audit cadence, verify Key Image placement, plant additions
mck-setup-payoffBuild the setup-payoff ledger; detect dangling setups and groundless payoffs
mck-specificity-forgeScan for generic language; forge concrete particulars from world bible
mck-voice-firstLock voice anchors before drafting; produce voice-anchors.md
mck-exposition-ammoConvert info dumps into "exposition as ammunition" — every fact fired in combat
mck-negation-of-negationDrive the central value to Corner 4 (corruption of the positive); test Crisis depth

V3 — Greatness Pursuit

SkillPurpose
mck-honestyHonesty Engine — TEST / STRESS / REPAIR the CI against the author's Truth Library; separates grounded ideas from bumper stickers
mck-surprise-plantSurprise architecture — DESIGN (misdirection plan), PLANT (dual-reading item placements), AUDIT (spawn surprise-auditor)
story-tournamentTournament generation — CONTROLLING-IDEA / INCITING-INCIDENT / PROTAGONIST / CRISIS / CLIMAX; enforces diversity, judges blind

Workflow Skills (/story-*)

Full lifecycle from seed to manuscript.

SkillPurpose
story-newStart a new project from any seed (dream, image, news, mood)
story-statusDashboard — what's locked, what's next
story-premiseGenerate and lock a Premise Card
story-spineBuild the story skeleton (Inciting Incident → Crisis → Climax)
story-castDesign the full cast as a system of pressures
story-actPlan an act's scene sequence
story-sceneDraft or revise a single scene with internal Beat Gate and rolling review
story-auditRun the full McKee critic suite
story-reviseMulti-pass revision orchestrator
story-publishFinal assembly and manuscript export

Refactored Skills (formerly agents)

SkillPurpose
arc-tracerPlot a character arc (iterative, in main context)
act-designerDesign act structure and rhythm
exposition-smugglerConvert info dumps into "exposition as ammunition"
key-image-curatorIdentify and thread the Key Image across the story
composition-conductorCross-scene craft audit (pacing, setups, image threading)
controlling-idea-architectForge or audit the CI — FORGE / STRESS-TEST / TRACE / REPAIR modes
wiki-librarianMcKee wiki INGEST / LINT / MIGRATE / REGEN (runs in main context; bilingual)

Infrastructure Skills

SkillPurpose
story-stop-lossStop-loss convergence protocol — iteration caps, Beat Gate diversity trigger at round 2, backtrack depth limit
story-beat-gateInternal Beat-level scan, blind critique, AUTO/REVIEW/REJECT classification, resumable ledger
story-writer-adjudicationTwo-stage blind A/B preference, reveal, and structured human finding decision
story-personaAuthor Persona — FORGE / LOAD / APPLY modes; decision filter for all aesthetic choices
story-tournamentTournament generation for high-stakes creative decisions (also listed under V3 methodology)

Agents

Drop into .claude/agents/ in your project directory.

Generator Agents (produce artifacts)

  • premise-prospector — 5-candidate premise slate
  • character-forger — Character File with True Character + Dimensions
  • structure-skeleton — Spine document
  • scene-architect — Scene Cards
  • beat-miner — Beat Sheets
  • prose-drafter — Long-form prose from beat sheets (new)

Critic Agents (audit, with fresh eyes)

  • cliche-hunter — Clichés vs. honored conventions
  • antagonism-stress-tester — Antagonism force balance
  • crisis-climax-auditor — Dilemma authenticity, climax causality
  • subtext-whisperer — Text ≠ subtext ≠ desire audit
  • continuity-supervisor — World rules, character knowledge, physics (new)
  • tournament-judge — Blind ranking of N candidates (new)
  • voice-drift-detector — Line-level voice consistency vs. anchors (V2)
  • specificity-auditor — Flags generic nouns/verbs; world-bible-aware (V2)
  • reader-simulator — Blind read; FULL-draft and WINDOW rolling engagement reports (V2)
  • pacing-analyst — Scene length, rhythm distribution, FULL-draft and WINDOW rolling pacing reports (V2)
  • blind-beat-critic — Beat-level blind critique with bounded scene context (new)
  • batch-beat-pattern-auditor — Prose-only cross-scene homogeneity audit before the writer decision (new)
  • diversity-challenger — Mechanism-level alternatives when a Beat repeats or will not converge (new)
  • surprise-auditor — Naive read + misdirection plan cross-reference; verifies dual-reading plants and re-read moment at Climax (V3)

Specialist Agents

  • cast-balancer — Cast pressure matrix
  • genre-cartographer — Genre Contract
  • setting-surveyor — Four-dimensional setting + world rules
  • wiki-librarian — McKee wiki maintenance

Installation

Quick install (recommended)

# Clone this repo
git clone https://github.com/Quentinbest/McKee-stroy-workflow.git

# Install skills globally + scaffold a project (one command)
cd McKee-stroy-workflow
bash install.sh --project-dir /path/to/your-story-project

The install script:

  1. Copies all skills to ~/.claude/skills/ (global, available in every Claude Code session)
  2. Copies all agents to your-project/.claude/agents/ (project-local, used by Agent tool)
  3. Scaffolds project directory structure from templates

Run bash install.sh --dry-run to preview without making changes.

Manual install

# Clone this repo
git clone https://github.com/Quentinbest/McKee-stroy-workflow.git

# Install skills globally
cp -r McKee-stroy-workflow/skills/* ~/.claude/skills/

# Install agents into your story project
cp -r McKee-stroy-workflow/agents/* your-story-project/.claude/agents/

# Scaffold a new project
mkdir -p drafts/my-story/{characters,scenes,prose,audit/beat-gate,audit/rolling}
cp McKee-stroy-workflow/templates/lifecycle.json drafts/my-story/lifecycle.json
cp McKee-stroy-workflow/templates/state.json drafts/my-story/state.json
cp McKee-stroy-workflow/templates/beat-gate-policy.json drafts/my-story/beat-gate-policy.json

Optional: verify

# Verify Beat Gate contracts and fixtures
node scripts/verify-beat-gate.mjs
node --test tests/*.test.mjs
node scripts/run-beat-gate-dry-run.mjs
node scripts/run-beat-gate-dogfood.mjs
node scripts/compare-beat-gate-critics.mjs \
  --output benchmarks/beat-gate-dogfood/isolated-comparison-2026-06-12
node scripts/run-writer-adjudication.mjs create \
  --input benchmarks/writer-adjudication/memory-tide-pilot.json \
  --output benchmarks/writer-adjudication/runs/2026-06-12-memory-tide-unresolved \
  --seed 20260612
node scripts/run-writer-adjudication.mjs create \
  --input benchmarks/writer-adjudication/glass-orchard-calibration-v2.json \
  --output benchmarks/writer-adjudication/runs/2026-06-12-glass-orchard-calibration-v2 \
  --seed 20260612-v2

For live unresolved critic findings, first join them to an explicit variant catalog with run-writer-adjudication.mjs prepare. The command fails on missing or duplicate beat_ref + predicate matches and reuses the same authority and calibration validation as hand-authored inputs.

The dry run creates a synthetic story project under a temporary directory and verifies deterministic cleanup, protected-field rejection, human decisions, non-convergence escalation, and rolling review artifacts without using a real manuscript.

The dogfood benchmark runs a four-scene synthetic Chinese story through the same runner, ledger, critic-fallback, diversity, and human-boundary contracts. The default fixture applies the writer's dated choice of ending A, commits the accepted batch, and writes rolling reports. Add --pending to reproduce the pre-decision AWAITING_WRITER boundary and verify that Premise, character desire, and final aesthetic judgment are not silently automated.

The critic comparison replays the original 12 candidate Beats through isolated scene critics and a separate prose-only batch pattern auditor. Its retained evidence shows that isolation helps local independence, but cross-scene homogeneity requires its own bounded audit before the writer decision.

Writer Adjudication Protocol V2 separates three decisions: blind prose preference, blind finding judgment, and source-aware variant disposition. Source roles remain hidden through Stage 2A. The runner requires explicit reconciliation before accepting a finding after meaningful_difference: no, records keep_baseline separately from adopt_challenger, and never applies a variant automatically.

The retained Memory Tide pilot completed two comparisons: both challengers won blind, both findings were accepted, and both revisions were adopted. Treat this as narrow prospective evidence, not a general critic accuracy claim.

The completed Glass Orchard calibration dataset V2 used the legacy combined Stage 2 protocol. It contains 12 comparisons across four synthetic scenes, including three sealed weak-challenger controls. Challengers won 9/12 overall, but two weak controls also won and control resistance was only 33.3%. That failure motivated Protocol V2's separate unsupported-finding controls and PASS / WARN / FAIL gates. The retained evidence remains unchanged and must not be treated as Protocol V2 validation.


Usage

Start a new story:

/story-new "I had a dream about a clockmaker who builds clocks that run backwards"

Work on a scene:

/story-scene 2.3

/story-scene now drafts candidate Beats, runs the internal Beat Gate, and asks for one consolidated Beat decision instead of one prompt per micro-patch.

Apply subtext discipline to dialogue:

/mck-subtext-5layer

Run the full critic suite:

/story-audit

The Architecture Decision

Skills are the spine; Agents are the bounded workers.

  • Skills run in main context — visible, iterative, composable, cheap
  • Agents run in isolation — for parallel work, adversarial critics, and long tasks that would burn parent context

See story-plan.md for the full design rationale, V1/V2/V3 roadmap, and the five hard problems the system addresses.


Based On

Robert McKee, Story: Substance, Structure, Style and the Principles of Screenwriting (1997).

The McKee wiki this system queries lives at LLM-Wiki-Story.

Ver o original no GitHub

Bom para

Relacionados