Self-Improve Workflow

Critical Rules

• •NEVER apply changes without explicit user approval
• •NEVER add language-specific rules to global CLAUDE.md unless the user confirms their stack
• •NEVER remove existing rules without explaining why
• •Prefer editing existing files over creating new ones
• •Keep CLAUDE.md under 200 lines — link to separate files for detail
• •Every rule must be trigger-action format: WHEN X, DO Y

Instructions

Step 1: Load Report

1. •IF $ARGUMENTS contains a file path, read that report file
2. •OTHERWISE, check for an existing report at ~/.claude/usage-data/report.html
3. •IF no report exists, tell the user: "No usage report found. Run /insights first to generate one, then re-run /self-improve." and stop.

Extract from the report:

• •Friction points and failure patterns
• •Interaction style analysis
• •What's working vs what's hindering
• •Specific suggestions from the report

Step 2: Audit Current Setup

Read these files to understand the current configuration:

• •~/.claude/CLAUDE.md (global rules)
• •~/.claude/settings.json (hooks, permissions, plugins)
• •Any project-level .claude/CLAUDE.md in the current working directory
• •All files in ~/.claude/hooks/
• •All files in ~/.claude/skills/

Compare the current setup against the report findings. Identify gaps: what problems from the report are NOT addressed by existing rules, hooks, or skills?

Step 3: Propose Improvements

Present a numbered list of proposed changes, grouped by type:

CLAUDE.md Rules:

• •New trigger-action rules that address identified friction
• •Rules to remove or update if they're stale or counterproductive
• •Keep rules language-agnostic unless working in a project with a specific stack

Hooks:

• •New hooks to enforce behaviors that rules alone can't guarantee
• •Changes to existing hooks that aren't working

Skills:

• •New skills to encode repeated multi-step workflows
• •Changes to existing skills

Workflow Changes:

• •Prompt patterns or habits to adopt
• •Session management improvements

For each proposed change, state:

• •WHAT: The exact change
• •WHY: Link to specific report finding
• •CONTENT: The exact text/code to write

Step 4: Apply with Confirmation

WAIT for user to approve, reject, or modify each proposed change.

• •Present changes one group at a time (CLAUDE.md, then hooks, then skills)
• •For each group, ask the user which items to apply
• •Apply only approved changes
• •After applying, summarize what was changed and what was skipped

Examples

Example 1: Post-insights improvement

User says: "I just ran /insights, now improve my setup"

Actions:

1. •Read report from ~/.claude/usage-data/report.html
2. •Audit current CLAUDE.md, hooks, and skills
3. •Present grouped proposals with rationale
4. •Apply approved changes

Result: Targeted config updates linked to actual usage patterns.

Example 2: Using a specific report

User says: "/self-improve ~/Downloads/report.html"

Actions:

1. •Read the specified report file
2. •Same audit and proposal flow

Result: Improvements based on the provided report.

Troubleshooting

No report found

Cause: User hasn't run /insights yet. Solution: Tell the user to run /insights first, then re-run this skill.

Report contains no actionable findings

Cause: Usage data is too sparse or setup is already well-optimized. Solution: Inform the user that no changes are recommended and explain why.