AgentSkillsCN

self-analyze

分析特定技能的执行轨迹，识别其中的规律、错误与低效之处。根据实际使用情况，提出改进后的SKILL.md草案。该草案将进入人工审核流程——不会自动生效任何更改。

SKILL.md
--- frontmatter
name: self-analyze
description: >
  Analyze execution traces for a specific skill to identify patterns, errors, and
  inefficiencies. Proposes an improved SKILL.md based on observed usage. The proposal
  is queued for human review — no changes are applied automatically.
license: Apache-2.0
compatibility: Requires SQLite storage and LLM access
metadata:
  tier: builtin
  mutating: "false"
  confirmation-required: "false"
  params: '{"skill_name": {"type": "string", "description": "The skill to analyze (e.g. web-fetch)"}, "window": {"type": "integer", "description": "Number of recent traces to analyze (default: 50)", "default": 50}}'

Instructions

Analyze recent execution traces for a named skill and propose SKILL.md improvements.

Parameters

  • •skill_name (string, required): Name of the skill to analyze
  • •window (integer, optional, default 50): How many recent traces to include in the analysis

Behavior

  1. •Query the execution_traces table for the most recent window records for skill_name
  2. •Compute statistics:
    • •Total invocations, success rate, average duration
    • •Most common parameters, most common errors
  3. •Retrieve the current SKILL.md body for the skill
  4. •Send traces + stats + current instructions to LLM with prompt: "Given these execution traces and current instructions, propose improvements to make this skill more accurate and reliable. Return: (1) proposed SKILL.md body, (2) rationale"
  5. •Insert the proposal into skill_refinements table with status = 'pending'
  6. •Confirm to user: "Proposal queued. Run /review to see and apply it."

Self-improvement cycle

This skill is the entry point to the self-improvement loop: self-analyze → skill_refinements (pending) → /review → accepted → SKILL.md updated

Example interactions

  • •"Improve the web-fetch skill based on recent usage" → skill_name: "web-fetch"
  • •"Analyze how memory-read has been performing" → skill_name: "memory-read"
  • •"self-analyze shell-exec" → skill_name: "shell-exec"