Improve Skill

Improve an existing skill based on conversation context and user feedback.

Inputs

• skill-name (required): Name of the skill to improve (e.g., "frontend-design", "pdf-editor")
• user-input (optional): Specific instructions on what to improve or focus on

Workflow

1. Parse Input

Extract the skill name and any additional user instructions from the command arguments.

If no skill name is provided, ask the user which skill they want to improve.

2. Locate the Skill

Search for the skill in common locations:

1. ~/.claude/skills/<skill-name>/SKILL.md
2. Project-local .claude/skills/<skill-name>/SKILL.md
3. If not found, ask the user for the skill's location

3. Read Current Skill

Load and understand the current skill's:

• SKILL.md content (frontmatter + body)
• Any bundled resources (scripts/, references/, assets/)

4. Analyze Conversation Context

Review the conversation history to identify:

Course corrections: Places where the user said things like:

• "No, I meant..."
• "That's not what I wanted"
• "Actually, do it this way instead"
• "You misunderstood"
• "Please change..."
• "That's wrong because..."

Repeated patterns: Things the user consistently:

• Asked for that weren't in the skill
• Had to clarify multiple times
• Preferred over the default approach

Explicit preferences: User statements about:

• Preferred formats, styles, or conventions
• Specific requirements for their workflow
• Things they always/never want

5. Synthesize Improvements

Based on analysis, determine what changes to make:

Frontmatter changes:

• Update description if trigger conditions should change
• Add missing use cases

Body changes:

• Add missing instructions or workflows
• Clarify ambiguous sections
• Add user-specific preferences
• Include examples that match user's actual usage
• Remove or modify instructions that led to incorrect behavior

Resource changes:

• Update scripts if they produced wrong output
• Add new references for missing domain knowledge
• Modify assets/templates to match preferences

6. Present Proposed Changes

Before making changes, show the user:

1. Summary of identified issues from conversation
2. Proposed changes to the skill
3. Ask for confirmation or additional feedback

Format:

## Identified Issues

Based on our conversation, I found these areas for improvement:

1. [Issue description] - [Where in skill this relates to]
2. ...

## Proposed Changes

### SKILL.md
- [Change 1]
- [Change 2]

### [Other files if applicable]
- [Change description]

Should I apply these changes? Any adjustments needed?

7. Apply Changes

After user confirmation:

1. Edit the SKILL.md file with improvements
2. Update any bundled resources as needed
3. Show a diff or summary of what changed

8. Suggest Testing

Recommend the user test the improved skill:

The skill has been updated. To test the improvements, try using
/<skill-name> with a similar request to see if it better matches
your expectations.

Guidelines

Preserve Skill Intent

• Don't fundamentally change what the skill does
• Add specificity without breaking generality for other use cases
• If changes would only apply to one user's context, consider noting them as "preferences" rather than hard rules

Be Conservative

• Only make changes clearly supported by the conversation
• When uncertain, ask the user rather than guess
• Prefer adding clarifying examples over rewriting instructions

Maintain Skill Quality

• Keep SKILL.md concise (under 500 lines)
• Follow progressive disclosure - move detailed content to references/
• Ensure frontmatter description remains comprehensive for triggering

Handle Edge Cases

No conversation context: If invoked without prior conversation about the skill:

• Ask user what they'd like to improve
• Offer to review the skill and suggest potential improvements

Skill not found:

• List available skills
• Ask user to confirm the skill name or provide path

User provides only general feedback (e.g., "make it better"):

• Ask clarifying questions about what specifically isn't working
• Offer to analyze the skill and suggest improvements

Example Usage

After a conversation where the user corrected output format multiple times:

User: /improve-skill frontend-design

Agent analyzes conversation, finds user repeatedly asked for Tailwind instead of vanilla CSS.

Agent response:

## Identified Issues

Based on our conversation, I found these areas for improvement:

1. CSS Framework - You corrected me 3 times to use Tailwind CSS instead
   of vanilla CSS. The skill doesn't specify a preferred CSS approach.

2. Component Structure - You preferred single-file components over
   splitting into multiple files.

## Proposed Changes

### SKILL.md
- Add instruction: "Default to Tailwind CSS for styling unless user
  specifies otherwise"
- Add instruction: "Prefer single-file components unless complexity
  requires splitting"

Should I apply these changes?

With specific user input:

User: /improve-skill pdf-editor focus on the rotation script, it keeps failing on large files

Agent reads the script, identifies the issue, and proposes a fix to handle large files better.