skill-evolution

스타13

포크2

업데이트2026년 5월 19일 10:33

Use when current skills/agents are insufficient, output quality was weak, user says a workflow failed, repeated errors happen, or a new reusable workflow should be added. Drafts improvements safely into skill inbox; never silently installs them.

설치

Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.

출처

allytag

allytag/claude_code

GitHub 저장소 열기 Creator 저장소 보기

다운로드

Skill Evolution

Use this skill to improve the LTS system without poisoning future work.

Safety Model

Never modify installed skills directly during drafting.
Draft new or updated skill into ~/.claude/skill-inbox/<skill-id>/SKILL.md.
Keep skill narrow: one workflow, clear trigger, minimal allowed tools.
Avoid large always-loaded text. Skills should be compact and specific.
Do not include secrets, absolute private project paths, or user API keys.
Do not include destructive commands, broad deletes, privilege escalation, or hidden network calls.
Run guard after drafting:

node ~/.claude/openrouter-claude-proxy/skill-guard.mjs validate ~/.claude/skill-inbox/<skill-id>

Promotion requires explicit user approval:

node ~/.claude/openrouter-claude-proxy/skill-guard.mjs promote ~/.claude/skill-inbox/<skill-id> --apply

Candidate Structure

~/.claude/skill-inbox/<skill-id>/
  SKILL.md
  notes.md        optional evidence: why this skill exists

Good Candidate Criteria

Clear name, description, and allowed-tools.
Description tells Claude exactly when to use it.
Tool list is least-privilege.
No overlap with an existing skill unless update is intentional.
Works for many projects, not one private folder.
Adds quality process, not shortcuts.

이 저장소의 다른 Skills

같은 저장소

agent-bootstrap

allytag/claude_code

Use when setting up or repairing repo agent instructions for Claude Code, Codex, or both. Proposes durable AGENTS.md, CLAUDE.md, and project memory changes before writing unless setup was explicitly requested.

2026-05-1913

api-contract

allytag/claude_code

Use for backend routes, API design, validation, database interactions, auth boundaries, response shapes, typed clients, error models, and integration-safe changes.

2026-05-1913

performance-pass

allytag/claude_code

Use for slow pages, large bundles, slow APIs, database performance, caching, rendering jank, long sessions, and production-readiness polish.

2026-05-1913

premium-ui

allytag/claude_code

Use for high-end UI, SaaS dashboards, landing pages, product websites, mobile responsive UI, visual redesign, UI polish, empty/loading/error states, motion, typography, and design quality rescue.

2026-05-1913

project-memory

allytag/claude_code

Use for long-running projects, context compaction, resume/handoff work, multi-session builds, or when the user asks to preserve project state. Proposes small verified memory docs before creating them unless setup was explicitly requested.

2026-05-1913

tool-coach

allytag/claude_code

Use automatically for Claude Code work involving files, tools, bash, edits, tests, long commands, failed tool calls, or unfamiliar repos. Improves tool choice, prevents loops, and keeps work high-quality without adding broad context.

2026-05-1913

name	Skill Evolution
description	Use when current skills/agents are insufficient, output quality was weak, user says a workflow failed, repeated errors happen, or a new reusable workflow should be added. Drafts improvements safely into skill inbox; never silently installs them.
allowed-tools	Read, Bash, Write

Skill Evolution

Use this skill to improve the LTS system without poisoning future work.

Safety Model

Never modify installed skills directly during drafting.
Draft new or updated skill into ~/.claude/skill-inbox/<skill-id>/SKILL.md.
Keep skill narrow: one workflow, clear trigger, minimal allowed tools.
Avoid large always-loaded text. Skills should be compact and specific.
Do not include secrets, absolute private project paths, or user API keys.
Do not include destructive commands, broad deletes, privilege escalation, or hidden network calls.
Run guard after drafting:

node ~/.claude/openrouter-claude-proxy/skill-guard.mjs validate ~/.claude/skill-inbox/<skill-id>

Promotion requires explicit user approval:

node ~/.claude/openrouter-claude-proxy/skill-guard.mjs promote ~/.claude/skill-inbox/<skill-id> --apply

Candidate Structure

~/.claude/skill-inbox/<skill-id>/
  SKILL.md
  notes.md        optional evidence: why this skill exists

Good Candidate Criteria

Clear name, description, and allowed-tools.
Description tells Claude exactly when to use it.
Tool list is least-privilege.
No overlap with an existing skill unless update is intentional.
Works for many projects, not one private folder.
Adds quality process, not shortcuts.