The era of informal, intuition-based coding and design just got a serious reality check. Bureau, an agent skill released on October 27, 2026, is forcing coding models to behave less like autocomplete engines and more like disciplined design studios. By imposing a rigid 10-gate review circuit, it aims to eliminate the generic, gradient-heavy output that plagues current AI-generated UIs.

The Tribunal Can Say No

Bureau operates on a simple, brutal premise: a model draws the average, so you need an editor to mark it. The skill introduces a Tribunal that reviews work in a fresh context. If a page fails any of the ten gatesβ€”covering everything from concept and type to accessibility and motionβ€”it gets sent back. It is not just about generating code; it is about enforcing a point of view. The system scores each gate from 0 to 5, and anything below 3.5 triggers a failure. It is a quality assurance loop that actually has teeth.

Killing the Generic Gradient

The tool's anti-slop detector is particularly ruthless. It flags common AI design clichΓ©s, such as the default purple/indigo gradient, glassmorphism by default, and the use of emoji as UI icons. In a proof-of-concept, a deliberately generic hero page was marked with 18 distinct flaws. These included three P0 critical issues: "Zoom disabled," "Gradient text," and "Lorem ipsum in shipped source." The detector also flagged "Native cursor hidden" as a P1 issue. This isn't just a linter; it's a taste engine that refuses to ship code that looks like it came off an assembly line.

One Skill, Four Hosts

Bureau is designed to be host-agnostic, working with Claude Code, Codex, Antigravity, and Hermes. The installation is straightforward: clone the repo and run the installer script, which copies the skill folder into the host's directory. The command structure is intuitive, using a /bureau router that loads only the necessary references for tasks like /bureau concept, /bureau build, or /bureau slop. It requires Node 20 or later, with Playwright and Figma support as optional, degrading gracefully if missing.

Key Takeaways

  • Bureau v7.10.0 introduces a 10-gate review circuit for AI coding agents, enforcing design discipline.
  • The Tribunal reviews work in a fresh context, ensuring objective quality control before launch.
  • The anti-slop detector specifically targets common AI design failures like generic gradients and hidden cursors.
  • The skill is compatible with major AI coding hosts including Claude Code, Codex, Antigravity, and Hermes.

The Bottom Line

Bureau proves that the bottleneck in AI-assisted design isn't generation speed, but editorial rigor. By hard-coding a refusal mechanism into the agent loop, it shifts the burden from the user's taste to the tool's constraints, finally treating AI output as a draft rather than a deliverable.