self-improve

Files (1)

• SKILL.md

SKILL.md (first lines)

# Self-Improve — Deep Optimization Cycle

## Trigger
`/self-improve`

## Purpose
Comprehensive self-audit and optimization. Read everything I have, find inconsistencies, gaps, and opportunities, then upgrade my own configuration.

## Workflow

### Step 1: Full State Audit
Read ALL configuration files:
- `~/.claude/CLAUDE.md` — global instructions
- All files in memory directory (`~/.claude/projects/*/memory/`)
- All skill files (`~/.claude/skills/*/SKILL.md`)
- All agent files (`~/.claude/agents/`)

### Step 2: Consistency Check
- Are there contradictions between CLAUDE.md and MEMORY.md?
- Are there duplicate instructions across files?
- Are there outdated facts that should be updated?
- Are there memory entries that reference completed work and can be archived?
- Are skills still relevant to the user's current projects?

### Step 3: CLAUDE.md Optimization
Based on the Anthropic Skills Guide and CLAUDE.md best practices:
- Is it under 50 lines? If not, consolidate.
- Are instructions actionable and specific? Rewrite vague ones.
- Are there missing instructions that past sessions suggest are needed?
- Run a search for latest CLAUDE.md best practices if significant changes are planned.

### Step 4: Memory Optimization
- Compress verbose entries without losing information
- Move stale project details to archive files
- Update progress sections to reflect current state
- Ensure cross-references between files are valid

### Step 5: Skill Gap Analysis
Based on conversation patterns in memory:
- Are there recurring tasks that don't have a skill yet?
- Are existing skills being triggered correctly?
- Do any skills need updated instructions based on experience?
- Should any skills be retired?

If gaps are found, either:
- Build a new skill (create folder + SKILL.md)
- Update an existing skill's SKILL.md

### Step 6: Report
Present a comprehensive improvement report:

```
## Self-Improvement Report

### Audit Results
- Files scanned: [count]
- Issues found: [count]
- Improvements made: [count]

### Changes Made
#### CLAUDE.md
- [change 1 and why]

#### Memory
- [change 1 and why]

#### Skills
- [new/modified/retired skills and why]

### Recommendations
- [things that need user input to decide]
- [suggested skills or tools to explore]

### Health Score
[Overall assessment of configuration quality: instruction clarity, memory freshness, skill coverage]
```

## Rules
- Always show changes BEFORE and AFTER so the user can verify
- Never delete information without asking — compress or archive instead
- If unsure about a change, list it under Recommendations instead of applying it
- The goal is QUALITY over QUANTITY — fewer, better instructions beat many vague ones
- Keep CLAUDE.md laser-focused on behaviour, not facts
- Keep MEMORY.md factual, not instructional

### HARD CONSTRAINT: Never degrade
- Self-improvement is monotonic — only ever improve, never regress
- Never delete a working instruction, memory, or skill unless replacing with something STRICTLY better
- Never weaken rules (e.g. "all systems max" must never become "moderate effort")
- Never remove user preferences or corrections — those are permanent
- Never overwrite files with shorter/vaguer versions to "simplify"
- When consolidating, the merged version MUST contain ALL information from originals
- If unsure whether a change is improvement or degradation — don't make it. Flag it for the user
- The ratchet only turns one way

From rafeeqinea/claude-academic-toolkit, MIT.

More from rafeeqinea/claude-academic-toolkit