Ask AI — generate tokens

Route: Library → Properties → Tokens → Primitives → Ask AI

The Ask-AI modal turns a prose description of a brand into a structured token patch — adds, updates, and removes you can review and apply.

Opening the modal

On the Primitives page, click Ask AI in the page header. The modal opens with a single textarea for your prompt.

When no API key is set, the modal shows a stub banner — the flow still works but returns a deterministic mock so you can test the diff/apply path without paying for a call.

Writing a useful prompt

Two patterns that work well:

Pattern A — brand brief

"Calm, professional palette for a fintech onboarding flow. Trust signals: navy primary, off-white surfaces, warm gold accent for the CTA. Avoid neon."

Pattern B — example anchor

"Match the feel of Linear's marketing site — muted neutrals, restrained accent, slightly warm grays in dark mode."

Both work because they give the model role (what the palette is for), mood (formal / playful / restrained), and at least one constraint (no neon, or 'restrained accent').

Reading the diff

The response is a TokenPatch:

• Add — new primitives the model proposed (e.g. color-navy-500, color-gold-400)
• Update — existing tokens whose value changed (shows old → new with swatches)
• Remove — tokens the model proposed deleting (rare; usually empty)
• Rationale — one paragraph explaining the choices

Each row has the proposed value with a swatch so you can scan the diff visually before applying.

Applying

Click Apply. The patch commits via the standard token-mutation actions (addToken / updateToken / deleteToken), which means:

• The change is on your undo stack — Ctrl/Cmd + Z reverts the whole patch.
• The Primitives page updates immediately.
• Every semantic and component binding downstream re-resolves.
• The Documentation tab regenerates fresh tokens.css, design-system.md, and design-system.json.

To dismiss without applying, click Cancel or hit Escape.

Stub vs real mode

Today the AI call surface ships with the stub wired and the real call deferred — see REBUILD_PLAN.md → M7. The diff/apply UX is final; the prompt that generates the patch is what's still landing.

Tips

• Don't ask for a full system — ask for primitives. Tostada's semantic and component layers stay yours to author.
• Iterate. Re-open the modal, refine the prompt, apply a new patch. Undo if you don't like it.
• Name your primitives. A patch arrives with names like color-navy-500. Rename them to your team's convention before binding semantics.