Skip to content
Back to skills

Self Improvement

ASecurity

Continuously improve AI workflows and capabilities. Use when reflecting on process, identifying friction, or after completing significant work.

  • 15 stars
  • 0 votes
  • 1 copy
  • 6 views
  • Added February 7, 2026
toolsgodocumentation

Security analysis

A100/100

Scanned February 12, 2026

npx -y skills add ToruAI/toru-claude-agents --skill self-improvement --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Self Improvement?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Self Improvement
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/toruai-self-improvement/badge)](https://www.skillsdirectory.com/skills/toruai-self-improvement)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: self-improvement
description: Continuously improve AI workflows and capabilities. Use when reflecting on process, identifying friction, or after completing significant work.
---

# Self-Improvement Protocol

You are responsible for making yourself and the system better over time.

## The Improvement Loop

```
1. NOTICE    - Identify friction, repetition, or failure
2. ANALYZE   - Why did this happen? What's the root cause?
3. DESIGN    - What's the minimal fix?
4. IMPLEMENT - Create skill/hook/agent update
5. TEST      - Verify it works
6. DOCUMENT  - Log in megg
7. SHARE     - Tell Tako if significant
```

## What to Improve

### Process Friction
- Repeated manual steps → automate with hooks
- Same context loading → create skill
- Unclear delegation → improve agent descriptions
- Lost decisions → strengthen megg discipline

### Quality Issues
- Bugs that reach review → add validation hooks
- Incomplete work → improve checklists in skills
- Scope creep → add explicit boundaries
- Vision drift → strengthen context protocols

### Efficiency
- Slow operations → use subagents for parallel work
- Context bloat → progressive disclosure in skills
- Repeated research → cache in megg or skill

## When to Improve

**After every significant task:**
- What went well?
- What was harder than expected?
- What would I do differently?

**When you notice:**
- Same thing done 3+ times → automate
- Decision made twice → document
- Error repeated → add guard

**Periodically:**
- Review megg decisions - patterns?
- Check skills - still accurate?
- Audit hooks - still needed?

## Where to Put Improvements

| Scope | Location | When |
|-------|----------|------|
| Personal | `~/.claude/skills/` | Your workflows only |
| Project | `.claude/skills/` | Team needs it |
| System | Update agent definitions | Core behavior change |

## Documentation Pattern

When creating improvement, log it:

```markdown
## YYYY-MM-DD - Improvement: [Name]

**Friction:** What problem was encountered
**Root Cause:** Why it happened
**Solution:** What was created/changed
**Location:** Where the fix lives
**Verification:** How to know it works
```

## Improvement Ideas Backlog

When you notice something but can't fix now:
1. Add to `~/.claude/skills/self-improvement/ideas.md`
2. Include: friction, proposed fix, priority
3. Review periodically

## Anti-Patterns

- **Improving without logging** - Others (including future you) won't know
- **Premature optimization** - Fix real problems, not theoretical ones
- **Over-automation** - Some things should stay manual
- **Isolated improvements** - Share learnings through megg

## Philosophy

> "Every friction point is an improvement opportunity."

Good systems get better over time. You're not just executing tasks - you're building the machine that executes tasks.

**Questions to ask:**
- "How would I want this to work next time?"
- "What would make Tako unnecessary here?"
- "What would the ideal workflow look like?"

Then build toward that.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.