skills/design-ai-production/references/production-system.md

# AI production system

## Stage contract

For every stage define:

- input assets and authoritative version;
- model/tool capability required, not just product name;
- prompt/template and variable fields;
- output format, dimensions, duration, and metadata;
- quality checks and measurable tolerances;
- retry limit and variation policy;
- fallback method;
- human approval requirement;
- cost and time assumption;
- provenance and license record.

## Consistency controls

- Canon and style bible
- Character/product reference sheets
- Fixed identity descriptors and negative constraints
- Approved palette, composition, motion, typography, and sound rules
- Seed/reference/version tracking where supported
- Reusable keyframes and source-of-truth assets
- Continuity ledger across batches
- Drift tests against approved anchors
- Final multimodal review after assembly, not only per asset

## Tool comparison fields

Capability fit, controllability, consistency, maximum output, speed, cost unit, expected retry rate, API/batch support, editing controls, localization, commercial terms, provenance features, regional availability, privacy, and fallback compatibility.

## Pilot

Test a representative small batch containing the hardest consistency case, one edge case, one localization case, and one complete final assembly. Measure acceptance rate, retries, cost per approved deliverable, cycle time, and dominant failure mode before scaling.