An evidence-backed Agent Skill that helps a coding agent design, implement, and review coherent standalone web applications with Fluent UI React v9. It focuses on application composition: which surface to use, where it belongs, and how the parts work together — not on repeated component API recall.

Scope

The skill answers composition questions:

Choosing a surface: a page, a drawer, a dialog, or an inline editor.
Arranging navigation, page chrome, commands, and content regions.
Deciding state ownership, the save model, and how submission reports back.
Laying out forms, data grids, tabs, and multi-step task flows.
Reviewing accessibility, responsive behavior, and visual coherence.

It does not restate component props, imports, or slots. Every rule and pattern is derived from a reviewed set of sources and cites the evidence it came from.

Division of labor with the sibling `fluentui` skill

This project is deliberately complementary to the API-focused fluentui skill produced by the sibling fluentui-mcp repository.

Skill	Owns	Use it for
`fluentui-design` (this project)	Composition rules and patterns, with sourced design evidence	"Where should this go, and how should the parts work together?"
`fluentui` (sibling)	Pinned API facts: exact props, types, imports, slots	"What is the exact prop or import for this component?"

The two skills are meant to be co-installed. fluentui-design links to fluentui for API facts instead of copying them, so API drift cannot silently desynchronize the two.

Install

Install the skill into the agent clients detected on your machine with one command:

npx -y fluentui-design skill install

The installer copies the bundled skill into each detected client's skill directory. Use the options below to change what it targets:

Option	Purpose
`--all`	Install into every detected client instead of asking.
`--target <dir>`	Install into a specific skill directory; repeat the flag for more than one.
`--project`	Prefer the current project's skill directory over the global one.
`--link`	Symlink the bundled skill instead of copying it.
`--dry-run`	Report what would change without writing anything.

Inspect or remove an installation with the status and uninstall subcommands:

npx -y fluentui-design skill status
npx -y fluentui-design skill uninstall --target <dir>

Usage

The skill is authored under skill/ and mirrored byte-for-byte to .agents/skills/fluentui-design/. Install it the way an agent discovers skills:

Copy .agents/skills/fluentui-design/ into your agent's skill directory, or point the agent at this repository's .agents/skills/ folder.
Install the sibling fluentui skill from fluentui-mcp next to it.
Start from skill/SKILL.md. It routes each task to the pattern or reference that resolves the decision.

The entry point lists the skill's triggers and non-triggers, and the generated reference index maps every decision to the patterns and rules that resolve it.

Version baseline

The bundled facts are pinned to one reviewed baseline. Update the table below deliberately (see MAINTENANCE.md); never edit generated output by hand.

Item	Value
UI library	`@fluentui/react-components` 9.74.7
React peer range	`>=16.14.0 <20.0.0`
API-fact source	`fluentui-mcp` enhanced schema pinned at commit d595d79

If an application uses a different package version, the skill tells the agent to say so and re-verify the rules it uses.

Repository layout

Path	Contents
`skill/`	The authored skill (`SKILL.md` + `references/`), mirrored to `.agents/skills/`.
`sources/`, `rules/`, `research/`, `facts/`	Source catalog, rules catalog, and analysis artifacts.
`scripts/`	Validators, generator, gates, and their tests.
`fixture/`	A Vite + React application that demonstrates one coherent composition, with Playwright + axe tests.
`evaluation/`	Fixed tasks, a rubric, and recorded deterministic evidence.

Commands

Command	Purpose
`npm run verify`	Full check: typecheck, lint, unit tests, catalog and skill gates, then Playwright + axe.
`npm run verify:static`	The browser-free subset; fast inner loop.
`npm run test`	Unit tests under `scripts/__tests__/`.
`npm run test:e2e`	Playwright tests under `fixture/e2e/`.

Known limitations

No model-provider evaluation. Agent-answer quality and the skill-versus-baseline comparison in evaluation/results.md are marked untested. The recorded evidence is deterministic: committed artifacts, gates, and accessibility findings, not generated answers.
Design assets are not fully analyzed. The official Figma kits and Teams UI templates (SRC-045) are recorded as a discovered expansion target; SRC-002 covers only the public getting-started page. Rules therefore lean on published guidance and the component library, not on Figma files.
Automated accessibility output is findings, not conformance. The fixture is scanned with axe and the results are reported as defects that were fixed; the project makes no WCAG conformance claim.

Maintenance

MAINTENANCE.md documents how to recheck sources, detect redirects and content changes, re-pin the baseline, find the rules affected by a fact change, rerun the fixture, and retain a history of replaced guidance. CHANGELOG.md records each skill version with its baseline. COMPLETION-REPORT.md closes the initial build honestly.