First read config.yaml(if it can't be read → state plainly that the configuration is missing and stop; don't use example values to pretend they're the user's settings), then read zmm/references/交互规范.md(🔴 Don't just read it once and consider it done: wrap up with the §4 trio — Recap · Before/After · next step gives numbered options; if information is missing, per §4 usemultiple-choice questionsto ask,only one at a time; for cases where it doesn't apply see §5), then read zmm/references/家族公约.md, then read memory {config.paths.memory}/zmm-skillify/ + _通用/。
This skill's built-in criteria are in references/规则卡.md(criteria / why / how to check / strength); read them once before starting;{vault} when there's a corresponding rule file in the vault, the vault takes precedence and the rule card is the fallback.
🔴 Gate: if it hasn't worked, it must not be solidified
This skill only distills from "a process that already produced a correct result." On entry, check three things first; if any is missing → state what's missing and stop:
| What to check | When not satisfied |
|---|---|
| Did something actually get finished in this session | Only discussion, no output → "there's nothing to solidify yet; finish this thing first" |
| Has the result been verified(user approval / data / an inspectable output) | Not verified → "this approach hasn't been proven effective; solidifying it is writing unverified experience as a rule" |
| Will it happen again | One-off → "this thing is unlikely to recur; the payoff from solidifying is negative" — you can just keep a record without making a skill |
⚠️ Don't skip the gate just because the user says "make it into a skill." What the user wants is to save effort next time, not one more inaccurate skill.
Where to distill from: three sources, ordered by credibility
- The actual process of this session — most credible. It was really done, there was output, there was feedback.
{config.paths.memory}skill memory in — corrections and effective methods accumulated before.- The user's spoken "how I usually do it" —— 🔴 Least credible. People's descriptions of their own process often differ from what they actually do. It can only serve as a clue; you must go back to ① or ② to find evidence; if you can't find it, mark it "unverified" in the final product.
When the three conflict, ① takes precedence, and write the conflict itself into the skill ("he says X, actually does Y, Y takes precedence" is often the most valuable rule).
Five steps
Step 1 · Write out the problem contract
反复出现的问题:
什么情境下会遇到:
这次是怎么解决的(按实际发生的顺序,不是理想顺序):
产出物是什么:
凭什么算做完了(可检查的证据,不是「感觉good」):
明确不处理的邻近问题:
🔴 "What counts as done" can't be written as checkable evidence → stop. A skill without a completion criterion means no one next time knows whether it ran correctly.
Step 2 · Translate "good to use" into observable behavior
Users will say "it should be written naturally" or "the judgment should be accurate."These cannot go into a skill, must be translated into something checkable:
必须做到(可观察):
禁止出现(可观察):
允许它自己发挥的部分:
最容易出的那个错:
出错时应该怎样(停下问 / 降级 / 明说做不到):
Contrast: "judges accurately" ❌ → "candidates must be chosen from the actually listed options; anything not listed must not be made up" ✅
Step 3 · Set credibility levels for the criteria
The family convention requires every conclusion to be labeled with its source.Every rule in a new skill must carry a level:
| Level | When to use | How to write it |
|---|---|---|
| 🟢 Measured | There is a specific case supporting it in this session or history | State the cost clearly: "Measured cost: full rewrite" |
| 🟡 Inferred | Derived from measurement, but not directly verified | Must add a sentence: "If this is wrong, where is it most likely to be wrong" |
| ⚪ Guessed | Just sounds right | By default, do not write it into the skill. If it must be written, label it firmly, and it cannot be used as a criterion |
🔴 Do not write 🟡 and ⚪ in the tone of 🟢. This is the most common mistake in the family.
Step 4 · Generate the skill directory and self-check
zmm-<名字>/
SKILL.md 入口:门禁 + 主流程 + 停止条件 + 红线
references/ 条件性细节(只在某些情况下才读的)
scripts/ 重复且确定的活(能用脚本就别让模型算)
Keep the main entry light: SKILL.md only holds what must be run every time; branch details go into references/。
Self-check list, go through each item:
- frontmatter has
name(ASCII slug, consistent with the directory name),description(includes Chinese trigger words),version - description clearly stateswhen it should be used, not just what it can do
- has a gate: under what circumstancesit should notbe used, and when to stop
- every rule carries a 🟢🟡⚪ level
- has a stopping condition: what counts as done
- has read the family convention and interaction spec, and ends with numbered options
- with the existing 20 skillsno overlap in responsibility — if there is overlap, modify that one, not create a new one
🔴 The last item is the easiest to miss. First run bash zmm/scripts/list-skills.sh to look at the existing ones, and explicitly state "what is the boundary between it and zmm-xxx ". If you cannot explain it clearly → you should not create a new one.
Step 5 · Let it run once before delivery (passing self-check does not mean it is usable)
The list above is all static items: files exist, fields are complete, rules have levels.If a skill is written and never run before entering the skill directory, you only discover next time it is actually used that the gate was written in a way that cannot run — this is the same discipline as "the person who writes the code cannot approve their own code". So before delivery, climb the ladder below, and at whichever level you reach, you may only say what that level allows:
| Level | What was done | What you are allowed to say |
|---|---|---|
| 0 | Static self-check passed | "The files are complete."You may not say "it works" |
| 1 | Take a real input, ina newly opened context , run it once according to SKILL.md; the output meets the completion evidence of the problem contract | "The main scenario runs through" |
| 2 | Ran a set of samples (see below), including at least onenear-neighbor counterexample, and the executor had not seen the answer | "It can distinguish what should be done from what should not" |
| 3 | After revising one version, rerun the same set of samples; the hit rate must not be lower than the previous version | "This version has not regressed" |
How to build samples(Put in the skills directory evals/, travels with the skill):
sample-NN-<描述>.md: input given to the executor.sample-NN-答案.md: given only to the grader; list two tables, "must catch" and "must not false-positive", 1 point per item- At least three types per group:Positive examples(should trigger and should be caught) · Boundary(right on the gate) · Near-neighbor counterexamples(look alike but belong to the neighboring skill, or look like a problem but are not) — the trigger words of the 22 skills overlap with each other; near-neighbor counterexamples are what prevent false triggers
- Record results in
evals/results/sample-NN-v<版本>.md,evals/README.mdas a table: sample · version · hits / full score · false positives · conclusion.The version number points to SKILL.md'sversion; do not call it "latest version" - Run by hand first; do not write a runner. Three samples run by hand for ten minutes is enough
If it fails, classify first, then fix(without classification you just keep adding rules to SKILL.md, making it longer and longer):
| Where it fails | What it looks like | Where to fix |
|---|---|---|
| Instruction | The rule is there; the executor understood it as something else | Change the wording of that rule; add a counterexample |
| Trigger | Should have entered but did not / should not have entered but did | Change the trigger words in description and "when not to use it" |
| Input | The sample itself is missing something; no one could do it | Change the sample, or add "stop if X is missing" to the gate |
| Criterion | The answer itself is disputed | First agree on the answer with the user, then touch the skill |
| Fluctuation | Same input gives different results twice | The rule gave room but no criterion; add a criterion |
🔴 If the same failure has gone through three revisions and you are still relying on "add one more rule targeting this sample", stop: that is not a missing rule; this skill's boundary is drawn wrong — go back to Step 1.
Ready-made examples:zmm-flow/evals/(run one version before and after the redesign; results for v0.1.0 and v0.2.0 are both there),zmm-resonate/evals/、zmm-review/evals/(created 2026-09-06).
Delivery boundary
By default, only make it work locally: generate the directory, put it into the user's skills directory, and tell them how to invoke it.
🔴 Publishing is a separate matter; do it only when the user explicitly asks. This includes: pushing to a public repository, publishing to ClawHub / SkillHub, adding to a packaged release. If the user did not say "publish it", do not touch these — the cost of changing a newly formed skill after it enters a public channel is much greater than keeping it local.
Stop conditions
Stop when any of the following is true; do not keep adding:
- The skill, ina new context, runs through the scenario in the problem contract, and the output meets the completion evidence (Step 5, level 1 or above)
- The user says it is enough
- You find it overlaps with an existing skill → modify that one, do not create a new one
- You find the gate was not passed → explain why; do not force it
Wrap-up
Per <Interaction Specification · IV>: Recap (what was solidified) · Before/After (which things would have to be rethought next time without this skill) · Next step given as numbered options.
Conveniently write a memory entry to {config.paths.memory}/zmm-skillify/: what was solidified this time, how the gate was passed, and whether any overlap was hit.Check for duplicates first。