discover
Understand a thing better by asking what it is an instance of outside this estate — handed to a separate reasoning process that has none of your context, so it answers in other people's vocabulary rather than yours. Use when about to say what something is, how it compares with its peers, or what sets it apart; when a shape is in front of you and your account of it is the only one you have; when about to generalise from one case; or when asked to look at something properly. Returns material, never a verdict.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- discover
- description
- Understand a thing better by asking what it is an instance of outside this estate — handed to a separate reasoning process that has none of your context, so it answers in other people's vocabulary rather than yours. Use when about to say what something is, how it compares with its peers, or what sets it apart; when a shape is in front of you and your account of it is the only one you have; when about to generalise from one case; or when asked to look at something properly. Returns material, never a verdict.
02 资源文件 1
- 参考文档EVIDENCE.md6.9 KB
03 SKILL.md 正文
<straw-dog question="q-uj5v4t">Mechanism: not yet</straw-dog>
Discover
You are the worst-placed party to say what a thing you built is an instance of. Not for want of knowledge — you hold a great deal — but because you are reaching for it while holding the frame that produced the thing, so you retrieve what confirms the shape you already have.
This hands the shape to a process with none of that, and files what comes back.
It returns material and never a verdict. What it names is a candidate for your case, not a finding about it. Why any of this: EVIDENCE.md.
What it is for, and what it is not
For: you hold one account of something and it is the only one you have. You are about to say what a thing is. You have one instance and are about to speak about the class.
Not for: ordinary recognition. Naming a thing correctly is not an error, and this does not fire on plain naming. It fires where the existence you are ascribing has quietly become one exhausted by your way of grasping it — where your handle on the thing has replaced it.
Never the reason to change something. A ladder always has a higher rung, so a pass will always return room for improvement; that is the shape of the question rather than information. Its output informs a change that already has a reason. A pass that motivates work has inverted, and the honest name for that is worthening.
Run it
Ask the question the shape answers, not the shape: what it is for, the situation it serves, without how this tree meets it — a shape described, in anyone's words, sends its frame along. No local file names, no local rule numbers, no local record names, no house terms. If the pass can only answer in your words, it never left — and this is the whole of the method, so it is worth more care than the rest of the tasking put together.
Transmit the frame as little as the subject allows, and say what you transmitted. The capture that matters is not the vocabulary of your files; it is the framing you cannot help sending, and a fresh process does not fix that by itself. A pass told what to look for will find it and the finding will be worth nothing.
Give it the rules it works under, written into the tasking — they are the pass's own instructions, not yours. Ask it for several families rather than one, each with where this sits on that family's own ladder, what the family already knows about how this goes wrong, and its confidence per claim.
Ask it to name what your tasking smuggled in. A first-class answer, and the one most likely to be the useful part.
Sort what comes back
Recency for evidence, longevity for principles. A claim that something works is evidence and must come from the same substrate, recent enough to be about it. A claim about how things of this shape behave in general is a principle, and there longevity qualifies — which is continued life, not elapsed time: still cited, still taught, still used to decide things. A dead old idea has age and no longevity. The same source flips value depending on which you asked it for, and nothing about the source changes.
Several doors, not one. A family is a full alternative approach to the thing, not a slice of it. So a disagreement between families is not a defect to resolve into one account — an answer that harmonises them has probably lost both.
Structural fit, not topical. Retrieval is cued by surface features, so an unconstrained what is this like returns subject-matter neighbours sharing no structure. Refuse a family whose mapping cannot be stated.
Stop where no further grounds appear. Depth is a consequence of that, never a parameter.
How to run it
Spawn a separate process — a general-purpose subagent — and tell it not to read this repository. Reading it re-contaminates, which is the one thing the separate process buys. Unless the shape is the repository itself: then the repository is the subject, and your description of it is the frame. The pass reads it cold, from a fresh clone, and gets only the question.
Never feed one pass into the next. A pass given an earlier pass's output retrieves what it already said, and the second isolation buys nothing.
Put the material into the work as it came back. A summary written before you use it is your
reading of the pass, which is the contamination the pass exists to escape. It goes into the body of
the question that prompted it, the way /impact's does, and to the person in the same turn: each
family in a sentence of their words, what it would mean here, and what the tasking was found to
smuggle — before any decision that stands on it is put to them.
Neighbours
The two passes are a pair, and the overlap is intentional. /impact is the first pass: what a shape is made of here, and what depends on it — it needs the corpus. This is the second: what the shape is an instance of elsewhere — it must not read the corpus. Running both means two processes with opposite permissions, which is the point of having two.
/align stress-tests a shape with a person in the room; this runs without one. A third pass, where a shape wants to go, is not built.
Output only:
Families
Each an alternative account of the shape — not a slice of one. For each: where this sits on that family's own ladder, what the family already knows about how this goes wrong, and confidence per claim.
What the tasking smuggled in
The framing the pass could see in how it was asked.
Nothing to take
Families refused for want of a statable mapping, and — when it applies — that the pass returned no usable material. That is a result, not a failed run.
04 同仓库更多 Skills
verify
Use after an implementation pass, or to verify a landed slice before close. Whole verification of landed work: ticket, RFC, governing docs, and the project's verification set.
ticket
Use to decompose a parent work into independently-workable child tickets. Use to mint ticket(s), possibly a single one, when planning implementation without one.
tdd
Test-driven development with red-green-refactor loop. Use when user wants to build features or fix bugs using TDD, mentions "red-green-refactor", wants integration tests, or asks for test-first development.
skill-up
Aid agent skill creation or modification. Use when creating a new skill or changing an existing one. Use whenever writing any agent instructions, whether to a file or directly in a reply.
setup-devops
Set up the project's engineering scaffolding, from toolchain to CICD and observability. Use when initializing a repo or filling a setup gap.
review-architecture
Explore the codebase for module-deepening opportunities — architectural improvements that raise testability and navigability. Use when asked to improve or refactor the architecture.