Bureau – BUREAU – an agent skill that designs like a studio

BUREAU v7.10.0, an MIT-licensed agent skill for Claude Code, Codex, Antigravity and Hermes released with a guide dated 2026-10-27, gives coding models a design-bureau workflow that runs from brief to launch through 10 review gates and a fresh-context "Tribunal." The skill ships 32 pattern rules in its slop detector and 21 project checks, scores each gate 0 to 5 with any gate below 3.5, any P0 or any missing evidence failing the release, and grades outcomes at 90+ Award Track, 82–89 Pass, 72–81 Conditional and under 72 Fail. BUREAU packages 11 commands — including init, concept, build, qa, slop, review and auto — in one skill folder across four hosts, and its author argues a model asked for a page with no direction "returns the centre of everything it has seen," which the review circuit is built to reject.

Job: BUREAU v7.10.0, MITLine guide of 2026-10-27Agent skill for Claude Code, Codex, Antigravity, Hermes Brief to launch,ten gates thatcan say no. BUREAU gives a coding model the workflow of a design bureau: concept, art direction, type, motion, accessibility, and a review circuit. Every page rides the same line, and the Tribunal can send it back. - pattern rules in the detector - 32 - project checks beside them - 21 - gates before a Tribunal - 10 - hosts, one skill folder - 4 Proof Every page gets a proof. The detector marks a generic generated page in red pencil, with the same rules the command line runs. Flip to the reviewed version, or proof this page. Argument A model draws the average. An editor marks it. Asked for a page with no direction, a model returns the centre of everything it has seen. That is not a prompt problem; it is a missing idea, a missing point of view and a missing review. Concept before pixels. Three directions that truly differ, each a blend of three designers on three axes, built from a survey of five verified references. No code before an idea, an emotional target and a world. Craft with limits you can check. Type, colour, composition and motion each have a reference and a budget; scripts measure what can be measured, from contrast and overflow to fingerprint divergence. A review that can refuse. Ten gates in order, evidence on every line, a ten-minute jury and a Tribunal in a fresh context. Invented proof, fake sends and hidden content fail whatever the score. Circuit Ten gates. One Tribunal. Each gate scores 0 to 5 against its weight; any gate under 3.5, any P0 or any missing evidence fails the release. Changing a source file makes every later gate stale. 90+ Award Track 82–89 Pass 72–81 Conditional < 72 Fail Judge before you measure. The design review is written before anyone reads the detector's output. The builder is the weakest reviewer. The Tribunal reads the work in a fresh context, or the run is declared degraded and cannot reach Award Track. Ten minutes for the jury. Ten to fifteen independent reviewers vote on every gate and every Law of UX; the best round inside the time box stands. Commands Call it by what you need. Type /bureau and a command, or ask in plain language; the router picks the command and loads only the references it needs. | Command | What it does | Leaves on disk | |---|---|---| | /bureau init | Writes the brief: audience, offer, sourced proof, non-goals. | BUREAU.md , config.json | | /bureau concept | Five-source survey, three directions, a blend for each, three comps. | survey.json , decision.md | | /bureau direction | Normative tokens, type specimen, colour in context. | DESIGN.md | | /bureau build | The real page, then seven refinement passes. | refinement.md | | /bureau motion | Five awarded motion references, then the grammar. | MOTION.md | | /bureau qa | Screenshots, contrast, overflow, fold and performance at five widths. | .bureau/qa/ | | /bureau slop | The detector you are watching, with fixes. | audit.json | | /bureau review | Ten gates, Delivery Gate, jury, Tribunal. | scorecard.html | | /bureau figma | Variables and frames to code, or an honest fallback. | tokens | | /bureau ship | Copy and claims check, metadata, launch list. | launch-checklist.md | | /bureau auto | All of it, stopping at three human checkpoints. | everything above | ShowCase Four studios that do not exist. Each was built with the skill from a blend of three designers and given a phone composition of its own. Open the ShowCase ↗ showcase/index.html Mountain observatory showcase/observatory/index.html Cerro Lumen Gerstner × Moholy-Nagy × Eames. Scroll from the dome to the galaxy, one power of ten at a time. Perfume house showcase/perfumery/index.html Maison Orvé Ruder × Greiman × Munari. Compose top, heart and base; watch them move through water. Ocean institute showcase/ocean/index.html Maré Sutnar × Wyman × Kare. Contours radiate from the name; raise the wind and the sea builds. Type foundry showcase/foundry/index.html Glifo A. Albers × Scher × Cooper. Pull five axes by hand, weave a field of letters, open one in depth. Install One folder, four hosts. Preview with a dry run, restart the agent, then call /bureau . Node 20 or later; Playwright and Figma are optional and degrade honestly. The installer copies the skill folder into the host's skills directory and marks it, so uninstalling removes only what it placed. Windows: installers/install.ps1 . Installation guide en/install.html . git clone https://github.com/acf1210/bureau.git cd bureau bash installers/install.sh --agent claude --scope user git clone https://github.com/acf1210/bureau.git cd bureau bash installers/install.sh --agent codex --scope user git clone https://github.com/acf1210/bureau.git cd bureau bash installers/install.sh --agent antigravity --scope user git clone https://github.com/acf1210/bureau.git cd bureau bash installers/install.sh --agent hermes --scope user Every host at once: bash installers/install.sh --agent all --scope user --dry-run Docs Read it in your language. Nine pages per locale, identical in structure: overview, quickstart, architecture, workflow, circuit, commands, anti-slop, visual language, install.