| name | update-skill |
| description | Use when updating existing skills based on conversation feedback, user corrections, or identified improvements. |
Update Skill
Update existing skills based on user feedback and corrections during conversation. This skill learns from the conversation to improve future outputs.
When to Use
Call this skill when:
- A skill-generated output required manual corrections during conversation
- User provided specific feedback on style, format, or content preferences
- Multiple iterations were needed to get the output right
- User expressed satisfaction after corrections and wants these preferences remembered
How to Use
After the conversation reaches a satisfactory state, invoke this skill:
/update-skill <skill-name>
Example:
/update-skill md-report-writer
When called, the AI will:
- Manually review the conversation for feedback and corrections
- Read the target skill file
- Identify which guidelines need updating based on the conversation
- Manually apply updates to the skill file
- Commit changes to git
What Gets Updated
The skill looks for these types of feedback in the conversation:
Style Corrections
- Tone changes (formal → casual, etc.)
- Formatting preferences (bullet points vs paragraphs, etc.)
- Emphasis changes (bold vs italics, etc.)
Content Structure
- Section ordering preferences
- What content should come first (narrative before data, etc.)
- Level of detail desired
Language-Specific Rules
- Quote mark preferences
- Punctuation conventions
- Terminology preferences
Code Presentation
- When to include full code vs snippets
- How much explanation code needs
- Comment style preferences
Citation and References
- When to cite papers/code
- How to format citations
- What level of documentation is required
Process
When you invoke /update-skill <skill-name>, the AI will:
-
Review the conversation to identify your feedback, corrections, and refinements
-
Read the target skill file to understand current guidelines
-
Identify what needs updating by matching your feedback to specific guidelines
-
Update the skill file manually based on what was learned
- Add new guidelines if needed
- Clarify ambiguous existing guidelines
- Add examples from the conversation if helpful
-
Commit to git with a descriptive message explaining what was learned
Example Scenarios
Scenario 1: User asked for more detailed code explanations
Conversation: User said "这个代码太复杂了,需要详细解释每一步的作用"
Action: Update the code explanation guidelines to require detailed explanations
for complex code, not just variable annotations.
Scenario 2: User preferred different quote marks
Conversation: User said "中文用中文的双引号!"
Action: Update the quote marks guideline to specify Chinese quotes for Chinese text.
Scenario 3: User wanted narrative before data
Conversation: User said "不要直接开始列表,先解释一下背景"
Action: Add/clarify the "narrative first" principle with stronger language.
Output Format
After updating, the skill reports:
- Which skill was updated
- What changes were made
- Git commit details
Example:
✓ Updated md-report-writer
- Added: Complex code requires detailed logic explanation
- Modified: Distinguished simple vs complex code handling
- Committed: a5e0e09
Limitations
- Only updates skills in
/Users/xjsongphy/.claude/skills/
- Cannot modify skill metadata (name, description)
- Requires user to explicitly call the skill
- Does not automatically update from every conversation
Best Practices
- Call when satisfied: Invoke this skill only after the output meets user expectations
- Be specific: The skill works best with concrete feedback patterns
- Review changes: The skill will show what it changed; verify it captured the right lesson
- Iterative refinement: Multiple conversations may be needed to fully capture a preference