skill-agent

name	skill-agent
description	Create, update, reuse, and diagnose Claude/OpenClaw-style skill packages
version	1.0.0
tags	["skills","lifecycle","emergence"]
triggers	["create a skill","update a skill","diagnose skill matching","reusable capability","建立 skill","创建 skill","创建能力","更新 skill"]
execution_contract	{"version":1,"min_pass_score":0.7,"rewrite_below_score":0.55,"disable_below_score":0.25,"min_samples_for_rewrite":3,"min_samples_for_disable":2,"criteria":[{"name":"transaction_outcome","description":"The request is resolved as created, updated, reused, queued, or failed with a concrete reason.","weight":0.4},{"name":"package_quality","description":"Created or updated skills use SKILL.md plus Markdown references, with scripts reserved for explicit helper code.","weight":0.3},{"name":"lifecycle_clarity","description":"The response and metadata explain lifecycle state and next availability.","weight":0.3}],"observable_signals":[{"name":"skill_creation_metadata","description":"Response metadata includes the skill creation status when a transaction occurred.","source":"execution_metadata","weight":0.3}],"failure_modes":["generic_advice_instead_of_transaction","duplicate_skill_created","missing_lifecycle_reason"]}

Instructions

Treat skills as first-class runtime objects:

1. Decide whether the user wants to create, update, reuse, or debug a skill.
2. Prefer standard packages with SKILL.md and references/*.md.
3. Reserve scripts/ for explicit helper scripts or source files, not prompt-readable Markdown checklists.
4. Explain lifecycle state clearly: created, updated, reused, disabled, or failed.
5. When debugging, inspect registry state, match reasons, lifecycle events, and logs.

Examples

Explicit skill creation

When the user asks to establish a reusable capability, create or update a skill package and return the lifecycle result instead of only giving advice.