Character Consistency

The single most visible failure in AI video: the same person looks like three different people across four shots. Models don't remember your character from shot to shot. Every generation starts from zero, so continuity has to be engineered with words, pictures, and IDs that don't change.

The workflow

1. Register everything that repeats. Every character, product, and key prop gets an asset ID before the first generation (references/asset-registry.md). The registry is the single source of truth.
2. Write the reference sheet. One canonical description per asset: physical, specific, frozen (references/reference-sheets.md). This text gets pasted verbatim into every prompt. Verbatim. Not "in your own words."
3. Attach reference images where the model allows. Image inputs lock identity far better than text. The adapters note per-model support.
4. Cite, don't describe. In every shot-brief, write the asset ID and paste the frozen description. Never paraphrase, never summarize, never "improve" it mid-project.
5. Verify across shots. Check identity at every join (../video-qa/references/continuity.md). Drift caught at the join is a reshoot; drift caught at delivery is a disaster.

Rules that don't bend

• One description, frozen. The moment two shots describe the same face differently, you have two faces. Copy-paste is a feature.
• IDs are contracts. mara-01 means the registry entry, not "a woman kind of like Mara." When the description needs to change, change the registry and bump the version. Don't silently edit prompts.
• Describe what's stable, not what's momentary. The reference sheet holds face, build, hair, wardrobe staples. Not expressions, not poses, not lighting. Momentary things live in the shot-brief.
• Products are characters too. A bottle, a car, a logo: same discipline. Geometry, colors, label placement: frozen and cited.
• Fewer recurring assets, better continuity. Every additional recurring character multiplies the drift surface. Design casts you can actually hold together.