AI can generate scenes.

It cannot maintain a world.
It cannot enforce continuity.
It cannot protect character psychology.
It cannot defend against plot holes.
It cannot manage cost.
It cannot scale without structure.

Narrative production is not prompting.
It is constraint management.

This framework exists because:

• Memory must be externalized.
• Quality must be gated.
• Costs must be bounded.
• Models must be interchangeable.
• Systems must survive scale.

If your workflow breaks at 30 scenes, it is not production-ready.

This repository is an attempt to make AI storytelling disciplined, composable, and durable.

AI Video Production Skills is a model-agnostic, production-grade framework for building AI-assisted films and series.

It provides:

• Character / Set / Voice anti-drift systems
• Truth validation with citation requirements
• Plausibility scoring frameworks
• World-state continuity tracking
• Narrative red-teaming
• Audience simulation personas
• Structured shot planning
• Prompt packaging standards
• QC grading systems
• Regen learning loops
• Cost routing and budget gates
• Tiered model escalation policies
• Full execution logging

This is not a prompt collection.

It is infrastructure.

Before generating anything, read:

👉 docs/how_to_run.md

That guide defines:

• Project structure
• Execution order
• Gate enforcement
• Loop behavior
• Lock discipline
• Budget controls
• Logging requirements
• Proxy → final escalation policy

Do not begin generation before reading it.

No reliance on chat history. All continuity is externalized.

Nothing progresses without passing validation checks.

The system is closed-loop, not linear.

Works with Claude, GPT, Gemini, MiniMax, Qwen, or any future model.

Budget gates prevent silent overspend.

Skills operate independently and integrate cleanly.

Everything is versioned.

• No artifact is overwritten.
• Logs are append-only.
• World-state updates require diffs.
• Lock updates require version increments.
• Generation must reference explicit artifacts.

If it is not in a versioned artifact, it does not exist.

Narrative production is structured across six layers:

1. Narrative Layer
   Story structure, character arcs.

2. Validation Layer
   Research, plausibility, red-team, audience simulation.

3. Continuity Layer
   Character locks, set locks, voice locks, world-state tracking.

4. Production Layer
   Shot planning, prompt packaging, proxy rendering, QC, regen learning.

5. Governance Layer
   Model routing and budget enforcement.

6. Observability Layer
   Run logging, cost tracking, regen tracking, audit history.

See docs/ARCHITECTURE.md for full system diagram.

Story
↓
Research → Plausibility → Continuity
↓
Locks
↓
Shot Plan
↓
GenPack
↓
Proxy Render
↓
QC
↓
Regen Learning
↓
Cost Optimization

Failure at any stage loops backward.

There is no forward-only mode.

See docs/loop_architecture.md.

This framework separates intelligence into tiers:

Throughput Tier

• Formatting
• Extraction
• Structured generation

Judgment Tier

• Red-team
• Arbitration
• Psychological validation

Research Tier

• Citation-heavy synthesis

Vision Tier

• Visual continuity validation
• Spatial reasoning

Always use the cheapest viable tier.
Escalate only when gates fail.

See docs/model_policy.md.

• story-architect
• script-breakdown
• character-psychology
• dialogue-consistency-check

• research-agent
• plausibility-auditor
• cultural-check
• narrative-red-team
• audience-simulator

• character-lock
• set-lock
• voice-lock
• world-state-tracker
• continuity-supervisor

• shot-planner
• prompt-packager
• storyboard-generator
• qc-grader
• regen-analyzer

• cost-optimizer
• model-router
• budget-gates

• run-logger

Each skill defines:

• Required inputs
• Required artifacts
• Pass / fail gates
• Escalation policy
• Loop integration
• Logging requirements

See:

examples/demo_short_film/

The demo includes:

• Versioned locks
• Truth packet
• Plausibility audit
• World state + state diffs
• Shot list
• GenPack
• QC report
• Regen rules
• Cost report
• Run logs

The example proves the system structure works end-to-end.

This framework assumes:

• 50+ scenes
• Hundreds of shots
• Multi-model routing
• Iterative QC loops
• Budget ceilings
• Long-running pipelines

If a system cannot scale without collapsing continuity, it is not production-ready.

• Not a screenplay template
• Not a one-click generator
• Not a vendor-locked pipeline
• Not a collection of “magic prompts”

This is production infrastructure.

1. Read docs/how_to_run.md (mandatory)
2. Review docs/ARCHITECTURE.md
3. Review docs/loop_architecture.md
4. Choose a data spine from docs/data_spine_templates.md
5. Initialize locks and world-state
6. Only then begin generation

Do not generate scenes before locks and validation loops are established.

• CLI orchestration starter
• Reference Google Sheets template
• Vision-based continuity validation
• Dialogue drift detection
• Editing heuristic skill
• Cost dashboard UI example
• Expanded example project

See CONTRIBUTING.md.

Contributions must:

• Define gates
• Define artifacts
• Remain model-agnostic
• Preserve loop integrity
• Respect budget policy
• Respect model policy
• Strengthen system discipline

MIT