AGENT SKILL

matrize-spread

Use when the design direction is genuinely still open and several deliberately divergent idioms should be compared side by side, each presented in one fixed shell so the comparison is of systems rather than of presentations. Trigger on 'show me a few different directions', 'four properly different options, not four shades of the same one', 'we haven't picked a direction yet', 'matrize spread'. Emits a forced-choice block and records the choice; until that record is answered the hook denies any write to tokens.json or out/. Not for variations inside one settled idiom — that is matrize-echo.

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-spread
description
Use when the design direction is genuinely still open and several deliberately divergent idioms should be compared side by side, each presented in one fixed shell so the comparison is of systems rather than of presentations. Trigger on 'show me a few different directions', 'four properly different options, not four shades of the same one', 'we haven't picked a direction yet', 'matrize spread'. Emits a forced-choice block and records the choice; until that record is answered the hook denies any write to tokens.json or out/. Not for variations inside one settled idiom — that is matrize-echo.
argument-hint
<system> [n]

03 SKILL.md 正文

Produce n deliberately divergent idioms and make choosing between them unavoidable.

Use this early, when direction is open. If direction is settled and only expression is open, this is the wrong skill — matrize-echo is the right one, and running spread there produces four answers to a question nobody asked.

Divergent means divergent

The failure mode is n variations of one idea wearing different colours. Before presenting, check each pair: if two proposals could be reconciled by changing one token, they are one proposal. Replace one.

Make them differ on something structural — density, the role colour plays, whether the system leads with type or with space, how much it asks the reader to infer. Each proposal states the premise it is built on, in one sentence, and premises that are not in genuine tension are not a spread.

Fixed shell, bounded specimen zone

The comparison must be of systems, not of presentations. Four divergent idioms rendered in four divergent shells cannot be compared at all — the reader is judging chrome.

So: one fixed shell for every proposal, with a bounded specimen zone per page rendered in that proposal's own idiom. The shell keeps the comparison honest; the specimen zone keeps each idiom judgeable on its own terms. Neither alone is enough — a purely fixed shell hides the thing being chosen.

(matrize-echo inherits the shell fully instead, because variations inside one idiom stay comparable regardless.)

The forced choice, and why it is enforced

Every spread ends with a block that names:

  • the decision required, in one sentence
  • what choosing each proposal commits to
  • what choosing each proposal forecloses
  • the cost of not choosing, stated concretely

Then write <root>/system/spread-choice.json with the proposal ids and chosen empty.

Until chosen is filled, the PreToolUse guard denies every write to <root>/system/tokens.json and <root>/out/. That is deliberate: an n-proposal portfolio with no forced choice is a procrastination machine, and a rule that depends on a model choosing to honour it is not a rule. Record the choice together with what is deliberately not adopted and why — that sentence is the part that survives into BRIEF.md.

Provenance still applies

A proposal is still derived. Each one cites the Design Cards behind it, and a proposal resting only on grade-C evidence says so on its own page rather than borrowing the confidence of the ones that do not.

If a proposal came from you rather than from a reference, the page says so. Opinionated defaults presented as derived findings are the one thing that makes the whole artefact untrustworthy — including the parts that were derived.

04 同仓库更多 Skills