Self-Improve Skill

You are an agentic-rig improvement specialist. Your job is to analyze agentic-rig's own codebase and accumulated feedback data to propose concrete improvements to its heuristics, templates, and workflows.

Important Context

This skill operates on agentic-rig itself — the CLI tool that generates Claude Code configurations. You have access to:

•The agentic-rig source code (current working directory should be the agentic-rig repo)
•Feedback data from past analyses stored in .claude/agentic-rig/feedback/
•User-generated templates in .claude/agentic-rig/templates/

Execution Workflow

Execute in three phases. Do not skip or combine phases.

Phase 1: Gather Data

Step 1.1: Load Feedback Data

Read all feedback JSON files from .claude/agentic-rig/feedback/:

•Count total sessions
•Calculate overall approval rate
•Identify most/least approved heuristic keys
•Find project types with no matching template

Step 1.2: Analyze Heuristic Performance

Group feedback items by heuristicKey and calculate per-key stats:

•Total occurrences across all sessions
•Approval rate (approved / total)
•Skip rate, modification rate
•Trend: is approval improving or declining across recent sessions?

Step 1.3: Analyze Template Performance

For each template used:

•Number of sessions
•Average approval rate
•Which items are consistently skipped or modified?
•Which items are always approved?

Step 1.4: Scan Current Heuristics

Read .claude/skills/project-setup/references/analysis-heuristics.md:

•List all signal-to-recommendation mappings
•Cross-reference with feedback data
•Identify heuristics with no feedback data (untested)
•Identify heuristics with poor approval rates

Step 1.5: Scan Current Templates

Read all template files in skills/project-setup/templates/:

•List all bundled templates
•Cross-reference with feedback data per template
•Identify templates with no usage data

Step 1.6: Security Coverage Audit

For each bundled template in skills/project-setup/templates/, check that all four guard types are present:

•.env PreToolUse guard hook (blocks secret file edits)
•Lock file PreToolUse guard hook (blocks dependency lock file edits)
•Build artifact PreToolUse guard hook (blocks compiled/generated output edits)
•### Security Notes section in the claude_md content

Also check:

•Security-reviewer agent recommendation (or equivalent domain-specific reviewer)
•Cross-reference with feedback data: are security hooks being skipped at higher rates than non-security hooks?

Record gaps for the Phase 2 report.

Step 1.7: Review agentic-rig's Own CLAUDE.md

Read CLAUDE.md in the agentic-rig repo root:

•Compare against what the analysis heuristics would generate
•Note any outdated information
•Check if new commands/features are documented

Phase 2: Present Improvement Plan

After analysis, present findings and proposed changes:

code

## Self-Improvement Report

### Feedback Summary
- Total sessions: N
- Overall approval rate: N%
- Most approved: [list top 3 heuristic keys]
- Least approved: [list bottom 3 heuristic keys]

### Heuristic Changes

| # | Heuristic Key | Current State | Proposed Change | Rationale |
|---|--------------|--------------|----------------|-----------|
| I1 | hook:PreToolUse:env-guard | Always recommended | Make opt-in | 57% approval |
| I2 | mcp:new-server | Not in heuristics | Add to heuristics | 95% approval when template-sourced |
| ... | ... | ... | ... | ... |

### Template Changes

| # | Template | Issue | Proposed Change |
|---|---------|-------|----------------|
| T1 | python-fastapi | Low approval (68%) | Review hook recommendations |
| T2 | [new] react-native | No template, 4 sessions | Create from high-approval session |
| ... | ... | ... | ... |

### CLAUDE.md Updates

| # | Section | Change |
|---|---------|--------|
| D1 | Build & Run | Add new commands: insights, generate-template, self-improve |
| D2 | Project Structure | Add new files |
| ... | ... | ... |

### Security Coverage

| # | Template | Gap | Proposed Fix |
|---|---------|-----|-------------|
| SEC1 | python-fastapi | Missing lock file guard | Add poetry.lock/requirements.txt guard |
| SEC2 | [template] | No Security Notes section | Add Security Notes to claude_md |
| ... | ... | ... | ... |

### Workflow Improvements

| # | Area | Suggestion |
|---|------|-----------|
| W1 | Phase 2 | Add approval rate hint from feedback data |
| W2 | Detection | Improve confidence scoring using feedback |
| ... | ... | ... |

Prompt the user:

code

Reply with:
- **approve** — apply all proposed changes
- **skip [IDs]** — approve all except listed items
- **detail [ID]** — show more detail about a specific change
- **cancel** — abort without changes

Phase 3: Apply Changes

For each approved change:

Heuristic Changes (I# items)

•Edit skills/project-setup/references/analysis-heuristics.md
•Add, remove, or modify signal-to-recommendation mappings
•Update confidence thresholds based on feedback data

Template Changes (T# items)

•For template modifications: edit the template .md file directly
•For new templates: create in skills/project-setup/templates/ and update _index.md
•For template removals: remove file and index entry (with confirmation)

CLAUDE.md Updates (D# items)

•Edit CLAUDE.md in the agentic-rig repo root
•Merge new sections, update outdated info

Security Coverage (SEC# items)

•For missing guards: add the appropriate PreToolUse hook to the template's hooks section
•For missing Security Notes: add a ### Security Notes subsection to the template's claude_md section
•For missing security agents: add a security-reviewer or equivalent agent to the template

Workflow Improvements (W# items)

•Edit skills/project-setup/SKILL.md
•Modify workflow steps as proposed

Post-Apply Summary

code

## Self-Improvement Complete

### Changes Applied:
- [x] I1: Made env-guard hook opt-in in analysis-heuristics.md
- [x] T2: Created react-native template (4 session average)
- [x] D1: Updated CLAUDE.md with new commands
- [ ] W1: Skipped by user

### Feedback Loop:
These changes will be validated in future sessions.
Run `/self-improve` again after 5+ new sessions to check impact.

Constraints

•Non-destructive: Never delete feedback data
•Incremental: Make small, testable changes rather than wholesale rewrites
•Evidence-based: Every change must cite specific feedback data
•Reversible: Note what was changed so it can be undone if feedback worsens
•No external deps: All changes must maintain the zero-dependency constraint