← 返回
未分类 中文

Standardize the de-upgrading and iteration of SKILL

Use when identifying skill candidates from repeated work, managing skill change proposals, reviewing candidates, or chaining skills into workflows. Also invo...
Use when identifying skill candidates from repeated work, managing skill change proposals, reviewing candidates, or chaining skills into workflows. Also invo...
chamberz40 chamberz40 来源
未分类 clawhub v1.0.2 2 版本 100000 Key: 无需
★ 1
Stars
📥 399
下载
💾 0
安装
2
版本
#latest

概述

Skill Spec Engineering

Overview

Skills are living specs that need continuous iteration. This methodology covers the full lifecycle: auto-detection → candidate logging → scaffolding → change management → downstream composition.

Phase 1: Detection

Hook-driven (0 tokens)

  • PostToolUse (async) — increments counter + logs tool names, pure shell
  • Stop — writes to data/candidates.md when threshold exceeded, injects one prompt (~30 tokens)

Default threshold: 15. Override: SKILL_CANDIDATE_THRESHOLD=20

Candidate format

Auto-written to data/candidates.md:

## 2026-05-03 | 23 calls | 6 tool types
- **Session:** abc123
- **Tools:** Bash,Edit,Write,Agent,WebFetch,Read
- **Status:** pending

Review mode

When invoked directly, read candidates.md and for each pending entry:

  1. Summarize key steps from the session
  2. Evaluate three criteria: multi-step, multi-turn, repeatable
  3. If YES → proceed to Phase 2 (Scaffold with duplicate check); if NO → mark skipped

Phase 2: Scaffold

Before creating a new skill, check for duplicates:

  1. Scan ~/.claude/skills/*/SKILL.md frontmatter (name + description)
  2. If a similar skill exists → create a proposal in that skill's CHANGE.md instead of a new skill
  3. If no match → generate from templates/SKILL.template.md
~/.claude/skills/[new-skill-name]/
├── SKILL.md        # Generated from template, pre-filled with steps
└── CHANGE.md       # Empty, ready for proposals

Phase 3: Change Management

Two tiers

TierCriteriaProcess
-------------------------
PatchWon't surprise users (wording, typo, edge case)Direct edit + git commit
ProposalChanges behavior (add/remove steps, reorder, change logic)CHANGE.md → user review

Rule of thumb: If someone using this skill would say "wait, what?" — it needs a proposal.

Proposal format

## #001 - [Title]
- **Status:** proposed | accepted | rejected | implemented
- **Date:** 2026-05-03
- **Trigger:** What scenario triggered this change
- **Proposal:** What to change + why
- **Breaking?:** yes/no

Phase 4: Composition

Next Steps format (append to each skill)

## Next Steps
- **If [specific condition]**: invoke [skill-name] — [one-line reason]
- **If [other condition]**: invoke [another-skill] — [reason]

Chains registry

data/chains.md records input/output relationships between skills:

[publisher-matcher] --{matched list}--> [publisher-review]
[publisher-review] --{confirmed list}--> [email-generator]
[email-generator] --{email content}--> [mail-send-batch]

After a skill completes, check chains.md for downstream → suggest to user.

Quick Reference

PhaseTriggerToken costOutput
-----------------------------------
DetectionEvery session (auto)0 (shell)data/candidates.md
ScaffoldUser confirms candidateOn-demandNew skill directory
ChangeIssue found during use0CHANGE.md or direct commit
CompositionAfter skill executionOn-demandNext Steps / chains.md

Common Mistakes

  • Treating patches as proposals — fixing a typo doesn't need a proposal, just commit
  • Letting candidates pile up — review periodically or they lose value
  • Vague chaining — Next Steps must include specific conditions, not "you might want to use X"
  • Treating scaffold as final — template output is a starting point, always customize

版本历史

共 2 个版本

  • v1.0.2 当前
    2026-05-08 13:14 安全 安全
  • v1.0.1
    2026-05-08 01:14

安全检测

腾讯云安全 (Keen)

安全,无风险
查看报告

腾讯云安全 (Sanbu)

安全,无风险
查看报告

🔗 相关推荐

ai-agent

self-improving agent

pskoett
记录自身发现以实现自我改进的技能
★ 4,124 📥 867,109
ai-agent

Skill Vetter

spclaudehome
AI智能体技能安全预审工具。安装ClawdHub、GitHub等来源技能前,检查风险信号、权限范围及可疑模式。
★ 1,243 📥 271,510
ai-agent

Self-Improving + Proactive Agent

ivangdavila
自我反思+自我批评+自我学习+自组织记忆。智能体评估自身工作、发现错误并持续改进。
★ 1,413 📥 325,418