AgentSkillsCN

skill-editing

当现有技能需要根据新证据进行更新时,应运用此技能。

SKILL.md
--- frontmatter
name: skill-editing
description: Use when an existing skill needs updating based on new evidence
trigger_pattern: When a skill receives negative feedback or new evidence contradicts it
tags:
  - meta
  - editing
  - evolution
version: 1
is_protected: true
q_value: 0.5
q_update_count: 0

Skill Editing

Iron Law

NO EDIT WITHOUT EVIDENCE.

Every skill edit must be justified by concrete evidence: feedback signals, new processes, or observed failures.

Decision Tree

code
Skill needs attention:
├── Q-value < 0.3 with 5+ uses?
│   ├── Fixable with content update?
│   │   └── YES → Edit content
│   └── Fundamentally wrong approach?
│       └── YES → Deprecate and create new
├── Q-value < 0.1 with 5+ uses?
│   └── Auto-suspended → Review for deprecation
├── New evidence contradicts skill?
│   ├── Minor correction needed?
│   │   └── Edit without version bump
│   └── Major change to approach?
│       └── Edit with version bump
├── Skill content is stale?
│   └── Update with version bump
└── Skill is protected?
    └── Requires force=True and clear justification

Edit Types

Minor Edit (no version bump)

  • Fixing typos or clarifying language
  • Adding small details that don't change the approach
  • Updating examples
code
skill_edit(name, new_content, version_bump=False, reason="Clarified step 3")

Major Edit (version bump)

  • Changing the approach or methodology
  • Adding/removing steps
  • Updating based on new evidence
  • Correcting wrong information
code
skill_edit(name, new_content, version_bump=True, reason="Added heap analysis step based on proc_xxx evidence")

Deprecation (for broken skills)

  • Skill is fundamentally wrong
  • Skill is superseded by a better one
  • Skill's domain is no longer relevant

Mark as suspended rather than deleting:

code
skill_edit(name, metadata_updates={"is_suspended": true}, reason="Superseded by skill-xyz")

Editing Protected Skills

Meta-skills (skill-discovery, skill-creation, etc.) are protected:

  • They require force=True to edit
  • Edits MUST be justified with clear evidence
  • The protection exists because meta-skills control the entire system

Evidence Requirements

ActionMinimum Evidence
Minor edit1 feedback signal or observation
Major edit2+ feedback signals or new process data
Deprecation3+ failure signals or clear supersession
Protected editDocumented reasoning + 2+ evidence points

Anti-patterns

  • Editing based on gut feeling without evidence
  • Making a skill more generic instead of more precise
  • Editing a skill that's working well (Q > 0.7)
  • Deprecating instead of fixing
  • Forgetting to provide a reason for the edit