Design director

Every task that touches a screen ends with a critic looking at it. The critic gets a screenshot and nothing else — no code, no diff, no previous critique, no round number — so the agent that built the screen is not the one deciding whether it worked.

What you get

Three tiers. The skill asks which one you want before it spends anything past the cheapest, and defaults to skipping when nobody is there to answer.

• Spot check — runs unasked at the end of any task that changed a surface: a pairwise comparison against the last screenshot, a rules tally against the brief, at most two nudges, no score. One call, about ten seconds.
• Review — for a surface whose baseline slipped. Three critic repeats per round on the current direction, scored, median reported. About two minutes per round.
• Redesign — for a new surface. A cached reference brief, three seeded directions each scored, then critic rounds until the score stops moving. The most calls of the three tiers by far.

Each design keeps a Markdown record on disk: the baseline per surface, every critic round with its median and spread, the Lighthouse scores, and the directions that were rejected with the model and date each was tried — so a direction that failed this month can be retried on a newer model.

How it works

In bb each critic is a hidden child thread spawned with bb thread spawn, with the screenshot attached as an image. scripts/bb-critics.sh handles the spawning, waiting and score parsing, and three repeats run at once. Critics prefer the claude-code provider when bb has one, then the session thread's own provider; DESIGN_DIRECTOR_PROVIDER overrides both.

The scoring critic is always an Opus-class model and the spot check never produces a number. Both come from a bench of eleven real sites: the smaller model picked the better of two screenshots 20 times out of 20, yet ranged from 1 to 7 scoring identical input in the same run.

Requirements

• Chrome or Chromium, on PATH or in /Applications, for the screenshots.
• Node 18 or newer for the screenshot script that drives Chrome over CDP.
• npx lighthouse for the accessibility gate, fetched on first use.
• Critic runs spend tokens on your own provider account. The cost depends on the model you pick and the tier: a spot check is one call, a redesign is dozens.
• Optional image and video generation keys (OPENAI_API_KEY, GEMINI_API_KEY, FAL_KEY) in a gitignored .env.agents. Without them the skill says so in one line and uses documented fallbacks.
• Only the Claude configuration is calibrated. The score findings above were measured on Claude models; a run on codex, pi or an ACP agent is a separate scale whose scores are internally comparable and nothing else, which is why the record logs the provider and model beside every score.

Credits

The framework comes from Anshu Chimala's How to turn your AI into a world-class designer; the tiers, the brief and the split score come from measuring it.