Drop a plain figma-links.txt (one URL per line) instead of pairing every image with a .figma.txt file — Claude Code pulls the screenshot/design context per link directly via the Figma integration. Also notes this same drop zone doubles as the intake for a future image-generation training set. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Reference examples (build-time inspiration)
Drop genuinely strong, on-brand OG/Thumb examples here so their design system can be extracted and encoded into the app. This is the brief's §6.8A / §12 flow.
Important: how examples actually inform the AI
The runtime suggester reasons over structured records (lib/ai/examples.ts →
FeaturedExample: subject → icon + template + pattern + why it works), not
raw pixels. So an image or Figma file doesn't feed the AI directly — it's
reference that a human (or Claude Code) reads to:
- tune the templates / design tokens (spacing ratios, icon placement, stroke & shadow by "mood", headline rhythm), and
- author structured
featured_examplesthat map a subject to a proven recipe.
Those structured records are what the AI uses.
What to drop here
Simplest: one text file of Figma links. No exporting, no pairing — just a plain list, one URL per line:
reference-examples/
figma-links.txt
https://www.figma.com/design/abc123/...?node-id=1-2
https://www.figma.com/design/abc123/...?node-id=3-4
https://www.figma.com/file/xyz789/...
Claude Code pulls the screenshot + design context for each link directly via the Figma integration — no manual export needed.
Alternative: paired image + note files, if you'd rather export PNGs yourself (§12):
reference-examples/
example-01.png # the rendered OG/thumb
example-01.figma.txt # the Figma URL (or export notes) for example-01
example-02.png
example-02.figma.txt
...
Either works, and they can be mixed. These files are internal-only (§11.6) and git-ignored — they inform the build; they're never shipped or exported.
Also used for: image-generation training reference
If we pursue a fine-tuned illustration generator (LoRA), this same drop zone is the intake for that training set too — the on-brand illustrations gathered here (via either format above) double as the reference set, so there's only one place to hand off images.
Then
Ask Claude Code to "analyze the reference-examples and update the featured
examples + templates." It will read the images (and any Figma links), extract
the patterns, and write structured entries into lib/ai/examples.ts (and tune
lib/design/{templates,tokens}.ts where the data supports it).