Skip to main content

Generation layer

The machine half of the brand OS: what you hand an image model to get on-brand output. Files live in static/brand/generation-layer/.

Fastest path: the hosted skill

Follow create-on-brand-image. It loads /brand.txt, picks the best image model in your harness, passes the Golden Atomic Brand References, honors the banned terms, and stamps the wordmark.

Master prompt

The always-on style preamble — copy it and prepend it to any asset request. Full text: master-prompt.md.

Master prompt (excerpt)
Visual asset for AI Performance Training (APT). Register: an elite training facility, an AI lab,
and a special forces program had a child. Should feel like Equinox, Anduril, Linear, Nike Training,
Whoop. Mood: athletic, rigorous, institutional, calm authority. Palette (dark is canonical):
near-black #050505 grounds, soft white #F3F3F3, exactly ONE accent — Exit Yellow #DFFF00 —
marking a single value moment. Flat and matte: NO shadows, NO gradients, NO neon, NO glow.
If photography: athlete ALONE, B&W, training facility / control room / empty gym — never smiling
teams or stock photos. If diagram/chart: outlined, geometric, minimal — no illustration.
Numbers in IBM Plex Mono. Headlines uppercase grotesk.

Example prompts

Per-asset starting points (athlete-alone hero, team scorecard, performance rail, output table, stat card): example-prompts.md. Each prepends the master prompt and names the reference GABR to pass in.

Reference set

Use the Golden Atomic Brand References as in-context image references. Always pass the matching reference for any reused element (the Lightning Bolt, a recurring figure) so style never drifts.

The always-on wordmark workflow

  1. Generate the art layer clean — reserve an empty band; don't ask the model to draw the wordmark text.
  2. Stamp the real wordmark PNG (/brand/logos/raster/wordmark-white.png or lockup-horizontal.png) into the band. Crisp and identical every time.

Banned terms

Visual + verbal bans, enforced on every render and every line of copy: banned-terms.md.

Visual spec

Per-asset canvas sizes, the two visual modes (photographic + vector data graphics), and rules: illustrations/SPEC.md.

The discipline

  • Store every render's prompt next to the output.
  • Validate each new exemplar with a human before it becomes golden.
  • Pass golden references back in for reused elements.
  • Exit Yellow only at the single value moment. No shadows. Numbers in mono.
  • The repo + BRAND.md stay canonical; this portal renders over them.