BUREAU 7.10.0 packages a design-review workflow for Claude Code, Codex, Antigravity, and Hermes. Its project page, dated October 27, 2026, describes an MIT-licensed skill that takes coding agents from a brief through concept, typography, composition, motion, accessibility, truth checks, and launch.
The central feature is permission to reject the work. Ten weighted gates require evidence, while a fresh-context Tribunal can send a page back for changes. The project also lists 32 detector pattern rules and 21 project checks.
That makes BUREAU an interesting answer to generic AI-generated interfaces. It separates recognizable source-code problems from decisions about whether a page has a convincing idea. The distinction matters: detecting a default purple gradient is straightforward; deciding whether a visual direction fits a product is editorial work.
What BUREAU Changes Before Code Starts
BUREAU begins by documenting the audience, offer, sourced proof, and non-goals in BUREAU.md. Its concept stage then requires a survey of five verified references and three distinct directions. Each direction combines three documented designers, with each contributing a different axis of the design. Direction work produces DESIGN.md, covering normative tokens, typography, color, and the visual world. These steps are described in the project’s workflow documentation.
The practical change is sequencing. A coding agent must explain the intended design before producing and polishing the page. That gives a reviewer something more specific to assess than whether the output looks professional.
I’d put more weight on this documented intent than on the detector’s ability to recognize fashionable effects. A gradient can serve a clear concept, and an interface without gradients can still be generic. A brief and an explicit direction create a basis for deciding which is which.
How the Ten Gates Control Release
BUREAU assigns each gate a score from zero to five and a weight. The weights total 100, with concept and direction together accounting for 29 points of emphasis.
| Gate | Review Area | Weight |
|---|---|---|
| G0 | Intake | 5 |
| G1 | Concept | 14 |
| G2 | Direction | 15 |
| G3 | Type and color | 12 |
| G4 | Composition | 12 |
| G5 | Motion | 12 |
| G6 | Engineering | 10 |
| G7 | Accessibility | 9 |
| G8 | Truth | 8 |
| G9 | Launch | 3 |
The published release rules prevent a strong total from concealing a critical failure:
- Any gate below 3.5 out of 5 fails the release.
- Any P0, the highest-severity finding, fails it.
- Missing required evidence also causes failure.
- Changing a source file makes subsequent gate evidence stale.
Aggregate scores have separate bands: 90 or above qualifies for Award Track, 82–89 is Pass, 72–81 is Conditional, and below 72 is Fail. Those labels remain subject to the blocking rules.
The stale-evidence rule is particularly useful. A page can pass accessibility review and then acquire a broken interaction during a final visual edit. Invalidating downstream review prevents an earlier approval from standing in for the changed page.
These numbers encode priorities, however. They are not a calibrated measurement of design quality. A concept score of four still depends on the reviewer’s interpretation of the rubric.
What the Detector Can Actually Check
BUREAU’s repository documentation describes its detector as the mechanical subset of the process: 32 regular-expression rules alongside project checks. The project page’s demonstration reports 18 findings, including three P0s, on a deliberately generic specimen created for that demonstration.
The sample includes:
- Disabled zoom, gradient text, and shipped lorem ipsum, marked P0.
- A default purple or indigo gradient and a default UI font used for display text.
transition: all, positivetabindex, anddivorspanelements used as buttons.- Generic marketing language and an unsourced credibility metric.
- Default glassmorphism and hover scaling applied to every card.
These checks mix different kinds of evidence. Finding transition: all in source code is a direct pattern match. Flagging a purple gradient identifies a stylistic choice, whose appropriateness depends on context. Detecting an unsourced number identifies something to investigate, not proof that the number is false.
That last distinction appears in the documentation itself. The demonstration marks an unsourced credibility metric as P1, while the repository treats fabricated metrics, testimonials, awards, and false submission-success messages as P0 problems. Missing attribution and fabrication therefore require different judgments. (Demo, repository)
The severity system leaves room for contextual decisions. P0 rules cannot be waived. A P1 waiver needs the rule ID and a page-specific reason of at least 20 characters. P2 findings must be addressed but never block release. That structure makes stylistic exceptions possible without silently ignoring a flag.
My view is that these rules work best as review prompts. A familiar visual pattern deserves scrutiny when an agent selects it automatically, but banning familiar patterns cannot by itself produce an original interface.
Concept and Direction Remain Editorial Decisions
BUREAU explicitly rejects the idea that a clean audit guarantees quality. Its repository says a page without a defined visual world can pass mechanical checks and still fail the concept and direction gates.
The project also instructs reviewers to write the design review before reading detector output:
“Judge before you measure.” — BUREAU project page
That ordering helps preserve the difference between evaluating a design and clearing a checklist. Otherwise, a reviewer could let easily counted defects define the entire assessment. (Source)
Some requirements are verifiable without being sufficient. A reviewer can confirm that five references were recorded, three directions were produced, and typography has documented tokens. Those artifacts do not prove that the references are relevant, the directions genuinely differ, or the type suits the product.
The same applies to motion. BUREAU calls for five awarded motion references and a documented motion grammar. That creates traceable intent. Whether animation clarifies an interaction, distracts from content, or feels appropriate remains a contextual decision.
For teams adopting the skill, I’d keep a human decision at the direction checkpoint. Three variations of the same centered hero can satisfy a numerical requirement while offering little real choice.
What Fresh-Context Review Adds
The Tribunal reviews the work in a fresh context and can return it for changes. BUREAU classifies review performed in the builder’s context as degraded and bars that run from Award Track. Its jury process calls for 10 to 15 reviewers voting in parallel within a ten-minute window, with the best round inside that window standing. ()
Frequently Asked Questions
3 questions
1Which Coding Agents Does BUREAU Support?
BUREAU 7.10.0 lists Claude Code, Codex, Antigravity, and Hermes as supported hosts. The skill is MIT-licensed.
