AGENT SKILL

matrize-survey

Use to scan the current state of the art in design-system delivery — token formats and tooling, build and bundler conventions per target, component distribution, the accessibility baseline, and how systems are actually documented today — because a design handbook does not go stale on style, it goes stale on the delivery chain. Trigger on 'what's the current state of the art for design tokens', 'is our delivery chain out of date', 'resurvey the tooling', 'matrize survey'. Every named package, tool, spec or API is verified to exist before the file is written; an unverifiable name is removed and logged, never softened into a hedge.

License 许可
MIT
所属仓库
Anselmoo/werkstoff
最近更新
2026-11-02
Stars
★ 1
Forks
0

01 前言 Front Matter

name
matrize-survey
description
Use to scan the current state of the art in design-system delivery — token formats and tooling, build and bundler conventions per target, component distribution, the accessibility baseline, and how systems are actually documented today — because a design handbook does not go stale on style, it goes stale on the delivery chain. Trigger on 'what's the current state of the art for design tokens', 'is our delivery chain out of date', 'resurvey the tooling', 'matrize survey'. Every named package, tool, spec or API is verified to exist before the file is written; an unverifiable name is removed and logged, never softened into a hedge.
argument-hint
[domain]

03 SKILL.md 正文

Answer the question a reference cannot: not what should this look like, but how are systems actually delivered right now. Re-runnable, because the answer decays.

What it covers

  • token format and tooling conventions — which spec revision is current, which tools claim support, and where that support is uneven rather than absent
  • build and bundler conventions per target in play
  • component distribution conventions
  • the current accessibility baseline, and what changed in it
  • how systems are documented today — which is not a 13-page landscape PDF, and the survey should say when that still wins anyway: it wins for approval by a non-technical decision-maker, and loses for developer consumption. Both halves are findings; recording only one is advocacy

The confabulation gate

Every named package, tool, specification or API is verified to exist before the file is written.

Verify by resolving the thing itself — a registry entry, the spec's own URL, the repository. Not by recalling it, and not by finding it mentioned somewhere.

A name that cannot be verified is removed and logged in a "could not verify" list. It is never softened into a hedge: "possibly some-tool" reads as a real option to the next reader, and a hedge is how a confabulated dependency survives review. Dispatch fact-checker for the existence checks — they are mechanical and verifiable by construction.

Per-claim volatility, and an expiry derived from it

Each claim carries: source URL, retrieval date, and a volatility flag — stable (a convention that moves on a multi-year cadence) or moving (tooling support, version coverage, anything mid-adoption).

The file's expiry is the earliest of its claims' horizons: roughly 365 days for a stable claim, roughly 90 for a moving one. A survey made entirely of stable conventions should not expire on a calendar that assumes churn; one resting on a single fast-moving tooling fact should not look fresh for a year.

matrize-status names which claim forces the expiry, so a re-survey can be targeted rather than total.

The delta section

If a previous STATE-OF-THE-ART.md exists, end with what changed since it: claims that held, claims that moved, and claims that turned out to be wrong. A survey with no delta against its predecessor has not been compared to it.

Worked example of why this phase exists

Uneven support for a stable specification is the characteristic shape of this domain: a spec can reach a stable revision while the major implementations are still catching up to it, so "tool X supports the spec" and "tool X supports the revision you are targeting" are different claims. That distinction was not true a year earlier and will not stay true — which is the whole argument for re-running this rather than writing it once.

Record the revision you actually target, per tool, with its retrieval date.

04 同仓库更多 Skills