/critique-committee
Run both critics on one target in parallel, then present their findings grouped by lens. The agents carry the critique criteria — your job is orchestration and presentation, so keep your own instructions minimal. The two axes — looks (design-critic) and use (product-critic) — stay separate; do not merge or rerank across them.
Steps
Resolve and verify the target once — the UI / flow to critique, who it's for, and how to reach the UI: the route/state and the launch command for the surface the app ships as, plus a screenshot for the visual lens if you already hold one. Verify the entry point loads real, authenticated data before you spawn, so each critic judges the running product rather than a blank shell. If you can't reach or authenticate it unattended, that's a loop-readiness gap — surface it (the
loop-check-readinessaudit diagnoses these) and stop. Resolve this once, then hand the same launch info to both critics and let each route to its own surface.Spawn both at once (one message, two Agent calls), foreground. Each critic owns its own capture and flow — the looks lens and the use lens deliberately look at different aspects, so let each grab the screens or walk the steps it needs rather than sharing one capture:
subagent_type: design-critic— the visual lens; give it the target + how to reach the UI (or a screenshot, if you already have one for it to judge).subagent_type: product-critic— the usability lens; give it the target + how to reach the UI.
When a UI is reachable and both will drive it, each critic captures independently — keep the two captures from colliding (each critic isolates its own session; the visual-loop skill owns the mechanism). This assumes the app can serve two drivers at once; if it can't without collision, surface that as a loop-readiness gap.
Report, grouped by lens. Present design-critic's findings and product-critic's findings separately — do not merge or rerank across the two. An empty findings array from either lens is a valid result; report it as such.
Before spawning, check both subagent_types resolve. If either is missing, stop before spawning
any agent and name the missing one — don't run a partial committee silently.