跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
浏览器与自动化

skillify

Distil a proven workflow from this conversation into a reusable Skill file.

Surething-io · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add Surething-io/cockpit --skill skillify

技能说明


name: skillify description: "Distil a proven workflow from this conversation into a reusable Skill file." argument-hint: "[placement directory] [target to skillify]"

Skillify

Abstract "the thing you just pulled off" into a reusable skill. This is an analyze → extract → save flow:

  • Step 1 is always analysis: first decide whether the conversation / recent context holds knowledge worth distilling into a skill. If not, say so and stop — never force it.
  • The placement directory and other args are only needed at the final save step — do NOT front-load a "where should it go?" question.

Applies when the user says: "abstract this into a skill", "write up what we just learned as a skill", "do this automatically next time", "skillify this", "capture this workflow".

Core goal: turn a one-time success into a stable procedure, not copy a transient chat into a prompt.

Arguments (all optional; used only at save time)

  • Trailing text (after /skillify) = the object / lead to skillify; if omitted, distil from recent context.
  • Placement directory <skills-dir> = where the skill lands; the canonical source goes to <skills-dir>/<slug>/SKILL.md. Only needed at the "save" step; ask only if the user hasn't provided it — never ask before analysis.

Step 1: Analyze — is there knowledge worth capturing (scan → gate)

Answer "should this even be extracted" first. Most candidates die here, not in the writing.

1a. Three-phase scan: don't miss invisible skills

Review what you just did, sweeping each phase once, asking one question per phase: is there a "non-obvious action / judgement" here?

PhaseWhat to scanWhy it's easily missed
DiscoveryHow did you locate / trigger / diagnose the real problem? Which signal did triage use?No visible deliverable — most often missed
DecisionWhat justified the judgement call at the sticking point? Any reusable criterion?Treated as "intuition", not realized it can be codified
SolutionIs the execution skeleton / tool-combo / verification reusable?Visible — easiest to fixate on

The scan is about breadth, anti-omission: people naturally notice only "solution" (it has deliverables), while "discovery / decision" — the diagnosing and judging moves — are high-value yet invisible.

The three phases are only for scanning candidates, not chapters in SKILL.md. The body should be organized by failure mode / investigation action, not by these three categories.

1b. The real gate: don't over-extract

For each candidate, all four must pass:

  • Recurs: will you hit this kind of scenario again? (One-off → a note is enough, no skill.)
  • Non-obvious: is there an action / judgement you "wouldn't know without looking it up"? (Obvious things aren't worth codifying.)
  • Cost: is the cost of getting it wrong / missing it high?
  • Stable: is the procedure stable, not bound to this session's transient context / data?

Missing "non-obvious" or "recurs" → basically drop it.

If nothing passes the gate → tell the user "not worth a skill this time", explain why, and stop. Do not force a deliverable, and do not ask for a placement directory at this step.

1c. Merge or split: the boundary question

If discovery / decision / solution all surface candidates in one scenario, decide whether they go into one closed-loop skill or split into a combination:

  • Will any single phase be invoked alone? Yes → split; always entered from the top → lean merge.
  • Is the value inside a phase or at the handoff? At the handoff (upstream expected-state / contract must reach downstream for verification) → merge; each phase self-contained, handoff carries no info → splittable.
  • Would merging make one unit carry orthogonal failure modes? Yes (e.g. static breadth vs dynamic-reasoning depth) → split.
  • Does this phase run first in other problems too? Fan-out ≥2–3 → split into a reusable leaf.

Third state (need both): when the loop must stay intact but a phase needs depth / reuse, use orchestrator + leaf — one closed-loop skill holds the contract and verification, delegating orthogonal phases to reusable sub-skills.

Default to one closed-loop skill. Split only on a real signal (a phase reused by a second scenario, or attention spread too thin for depth).

1d. Classify: decides how to write, not whether to extract

TypeExampleSkill focus
Investigationpull sources → build timeline → locate issueevidence sources, steps, output template
Debugreproduce → red test → fix → verifyinvariants, tests, acceptance
Tool-combochain multiple log/monitor/analysis toolsquery order, correlation IDs
Habit/conventiondirectory structure, naming, placementpath rules, naming rules
Writing templateissue / PR / retro reportstructure, tone, prohibitions

Do NOT write one-off details into the skill (a transient session, a full specific chat, a temporary verification code, etc.).

Step 2: Extract — distil stable principles and draft

If 1c decided on a split or orchestrator + leaf, run Step 2 once for each skill to land (each its own directory).

2a. Distil stable principles

Break the experience into:

  • Trigger: when this skill should be used
  • Input: what the user might provide, what to ask when it's missing
  • Evidence sources / tools: which sources, files, CLIs, logs to check
  • Steps: the stable, reusable execution order
  • Output: what to hand the user, or what artifact to create
  • Boundaries: what NOT to do, when to stop / ask
  • Verification: how to confirm the skill worked

Write principles, not a play-by-play.

2b. Choose a slug

Slug: lowercase, alphanumeric + hyphen, short and clear, verb/task- or topic-oriented. Before drafting, ls <skills-dir>/ to avoid collisions and match existing naming style (when the save di

阅读完整来源 ↗