reference-builder (formerly shot-builder)

What it does. Makes the reference images a film needs before any video is attempted: 3-view character turnarounds, 2x2 location plates, style sheets, single stills, 3x3 and 6-panel contact sheets, cinematic plates across five camera modes. Each locked cast, location or style is saved as a reusable <TOKEN> whose prompt block is pasted verbatim into every later shot, so identity, wardrobe and place stop drifting between generations.

Why it exists. Video models are only as consistent as the references they are handed. This skill builds and files those references; action-builder and mograph-builder consume them.

Runs the generation. FAL.ai, Kie.ai, Arcads (MCP) or prompt-only (writes the prompt to a file for Midjourney, ComfyUI, Higgsfield or any UI — zero spend).

Input	the brief from brief-builder, face and location photos, style references
Output	images in <output_dir>/<date>/<project>/ with a .md sidecar (prompt, cost, tokens); tokens in workspace/
Feeds	action-builder (reference-to-video), mograph-builder (style sheets)
Spends	per image on FAL/Kie; nothing in prompt-only mode

Files

• SKILL.md — hard rules, setup gate, token workflow, the five camera modes, video kickoff
• references/ — character-sheet, defaults, onboarding, cinema grammar
• providers/fal.md, kie.md, arcads.md — model ids, prices, quirks per provider
• scripts/setup.py (first run), project.py, token_resolver.py, fal_gen.py, kie_gen.py, prompt_only.py, cost.py, _common.py
• workspace/ — ships EMPTY (cast/ locations/ styles/ projects/); the skill fills it per user

Setup

python scripts/setup.py asks seven questions (provider, key location, default model, resolution, aspect, output folder). The default output folder written into setup.py and references/defaults.md is the author's own path — accept the prompt's suggestion or type your own on first run. Keys go in ~/.frmwrkd/.env. Python 3, standard library only.