Product Designer
Produce a screen that matches the design system it was given — not the model’s own house style.
-
1 Read the system Claude Haiku 4.5 480ms
14 colour tokens, 7 spacing steps, 9 components. Accent is silver — this system has no brand hue.
-
2 Pick the archetype Claude Opus 4.8 1120ms
A catalog is browsed, so cards — not a table. The system’s own rule against card grids applies to record lists in consoles, not galleries.
-
3 Build the screen Claude Opus 4.8 1670ms
Composed from existing components. Wanted a purple CTA out of habit; the system says silver, so silver.
Each step ran on whichever model is best at that job — not one model for everything. The worker is the workflow.
Permissions & guardrails- Read: design-system tokens, the component inventory and existing screens for the surface it’s building
- Write: application code only; it composes from published components and never edits the design system itself
- Escalate: the designer when the system has no component for the job, or when two tokens could both apply
- Never: invent a token, colour or spacing value that isn’t already in the system
- Never: fall back to its own house style when the system is silent — it asks instead
- Catalog screen built from 9 existing components. Zero new tokens.
- Rejected its own default palette twice — the system won both times.
- 1 escalation: no component exists for the accuracy bar. Flagged, not improvised.
81.6%
a general model gets 64.8% · +16.8pp
64.7 · 69.9 · 74.1 · 77.6 · 80 · 81.6
private rules were taught in one workspace and never leave it. general rules are true about the work, so every copy of this worker gets them.