research-grounded-world-design

Internal world-design stage behind the storyboard entry point. Builds reusable visual assets — scenes, characters, costumes, props, architecture, atmosphere, continuity — from a script, confirmed prod

SKILL.md

GitHub

Research-Grounded World Design

Core Contract

One reusable visual-world package before shot production. Every selected subject gets a stable ID, evidence class, written specification, prompt, reference binding, generated asset, and continuity key. One complete subject view per image, registered immediately with absolute paths.

Production depth controls coverage. Card completeness, source or world-rule binding, continuity keys, image inspection, and manifest traceability stay constant at every depth.

Capabilities

Load imagegen before the first generation call, and web-access before researching real subjects. Without imagegen, produce cards and prompts and stop before §6. Without web-access, work from local material and mark gaps 待核实.

0. Load The Confirmed Config

When called from the storyboard workflow, read planning/production_config.json and carry across: evidence_mode, production_depth, world_design_scope, image_generation_scope, bible_sections, quality_floor, script path, aspect ratio, visual style, reference roots.

Depth Entity coverage View coverage
compact core recurring entities one identity-defining view, plus continuity-critical views
standard all recurring entities identity, scale, handling, and location-continuity views
detailed all visible entities plus action phases, material detail, and state changes

Gate C-W: the world run records the same evidence mode, depth, scope, and quality floor as the storyboard config.

1. Resolve Evidence Mode

Read evidence-modes.md. Select strict_evidence, hybrid_evidence, or fiction_consistency, then classify each entity as historical, contemporary_real, scientific, invented, or stylized. A real entity inside an invented story activates hybrid_evidence for that entity set.

2. Start An Isolated World Run

python3 scripts/init_world_run.py \
  --project <project-dir> --slug <short-name> \
  --evidence-mode <mode> --production-depth <depth> \
  --world-design-scope <core_recurring|all_recurring|all_visible> \
  --bible-section world,characters,props \
  --production-config <storyboard-run>/planning/production_config.json

Declare only the bible sections this project contains; style and continuity are always created. Read world-design-contract.md before writing cards or manifests.

3. Register Evidence Or World Rules

Real entities: read local sources, prior fact-check records, approved photographs, maps, and documents. Register direct sources in research/source_register.csv, save usable images to research/reference_images/, and bind each entity to source IDs, observable facts, rights notes, and local paths.

Invented worlds: define physics, technology, social structure, geography, material culture, scale, palette, and motifs in research/world_consistency_brief.md and research/evidence_profile.json.

Keep style references separate from factual evidence in every mode.

4. Build Design Cards

One card per reusable entity in design/design_cards.json, typed as world, location, character, costume, prop, architecture, atmosphere, vehicle, creature, material, graphic, or typography. Each carries a stable ID, evidence class, source bindings, visible specification, continuity keys, prompt, required views, generated asset IDs, approval status, depth, and selection reason.

5. Build The Visual Bible

Write positive visual specifications into the declared bible/<section>.md files. Define recurring relationships explicitly: character scale within locations, ownership and handling, family resemblance, screen direction, light direction, weather progression, state changes, palette hierarchy.

Gate W: every entity selected by world_design_scope has one approved card and at least one approved reference or generated image.

6. Generate Reference Assets

Generate each declared view as one complete image — one subject, one view per file:

generated/<asset-type>/<design-id>__<view-slot>__v001.png

Append each result to manifests/assets.json with design ID, view slot, version, prompt, evidence IDs, generator, absolute path, created time, and approval status. Replacements increment the version; earlier versions move to garbage_pile/asset_versions/.

7. Validate And Deliver

python3 scripts/validate_world_manifest.py --run <world-run-dir>

Inspect each approved image against its card: real entities against bound evidence (silhouette, proportions, construction, material, scale, usage), invented entities against declared rules and continuity keys.

Export to deliverables/: <title>_visual_bible.xlsx, <title>_design_cards.json, <title>_asset_index.csv, <title>_prompt_book.md, <title>_world_design_qa.md.

Storyboard Handoff

Hand back the absolute world-run path, design/design_cards.json, approved bible/ files, the approved asset manifest, required design IDs, evidence mode and entity classes, and continuity keys. Storyboard shots reference these through world_asset_ids. Reuse the same design IDs across episodes; a changed specification creates a new asset version.

Completion Report

Depth, scope, evidence mode, source count, selected entity count, card count by type, generated and approved counts, replacements, fidelity or consistency issues, quality-floor status, delivery paths.