| 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.
