---

name: meta-iterate description: Analyzes session performance and iterates agent/skill/rule prompts for self-improvement, and provides skill development guidance. This skill should be used when user says "improve prompts", "analyze sessions", "self-improve", "discover skills", "compound learnings", "learn from sessions", "优化工作流", "迭代agent", "分析session", "发现新skill", "积累经验", "学习总结", "提取规则", or when developing skills with "write skill", "create skill", "开发 skill", "插件开发", "skill development". Triggers on /meta-iterate, 自我改进, 会话分析, prompt优化, skill开发. model: opus memory: user context: fork allowed-tools: [Read, Glob, Grep, Write, Edit, Bash, WebSearch, Task, TaskCreate, TaskUpdate, TaskList, AskUserQuestion]

meta-iterate - Self-Improvement & Skill Development

Analyze Claude Code session performance, iterate on prompts, and guide skill/plugin development.

When to Use

Self-Improvement Mode

• Periodically evaluate workflow effectiveness
• After noticing repeated issues
• When prompted by session-end reminder (every 10 sessions)
• To proactively improve Claude capabilities
• To compound learnings into permanent artifacts

Skill Development Mode

• Creating new skills
• Improving existing skills
• Developing plugins
• Auditing skill quality

Commands

Self-Improvement Commands

Command	Purpose
/meta-iterate	Run full 5-phase workflow
/meta-iterate evaluate	Only evaluate sessions
/meta-iterate discover	Discover new skill opportunities
/meta-iterate compound	Transform learnings into skills/rules
/meta-iterate diagnose	Only diagnose issues
/meta-iterate propose	Only generate proposals
/meta-iterate apply	Apply approved changes
/meta-iterate verify	Verify improvement effects

Skill Development Commands

Command	Purpose
/meta-iterate skill-create	Create new skill from template
/meta-iterate skill-audit	Audit existing skill quality
/meta-iterate skill-improve	Improve specific skill

Workflows

Self-Improvement Workflow

evaluate → discover/compound (optional) → diagnose → propose → [approve] → apply → verify

Phase	Input	Output	Agent
evaluate	Braintrust logs	EVAL-<date>.json	evaluate-agent
discover	Evaluation data	DISCOVER-<date>.md	(built-in)
compound	Learnings files	Artifacts proposal	(built-in)
diagnose	Evaluation	DIAG-<date>.md	diagnose-agent
propose	Diagnosis	PROP-<date>.md	propose-agent
apply	Proposals + approval	Component files	apply-agent
verify	Post-change sessions	ITER-NNN.md	verify-agent

Skill Development Workflow

design → create → validate → improve

Phase	Input	Output
Design	Skill concept	Design decisions
Create	Design	SKILL.md + references/
Validate	Skill files	Checklist results
Improve	Audit findings	Updated skill

Options

Option	Default	Description
--recent N	10	Number of sessions to analyze
--target PATH	all	Specific component to focus on
--type TYPE	all	agent, skill, rule, or all
--threshold N	70	Score threshold for recommendations

Examples

# Full workflow
/meta-iterate

# Evaluate recent 20 sessions
/meta-iterate evaluate --recent 20

# Discover new skill opportunities
/meta-iterate discover

# Compound learnings into permanent artifacts
/meta-iterate compound

# Focus on specific agent
/meta-iterate --target agents/plan-agent.md

# Apply specific proposals
/meta-iterate apply --proposals PROP-2026-01-10.md

# Verify after improvements
/meta-iterate verify

Output Files

Phase	Location	Format
Evaluate	thoughts/evaluations/EVAL-YYYY-MM-DD.json	JSON
Discover	thoughts/discoveries/DISCOVER-YYYY-MM-DD.md	Markdown
Compound	thoughts/proposals/COMPOUND-YYYY-MM-DD.md	Markdown
Diagnose	thoughts/diagnoses/DIAG-YYYY-MM-DD.md	Markdown
Propose	thoughts/proposals/PROP-YYYY-MM-DD.md	Markdown
Apply	Component files + thoughts/iterations/ITER-NNN.md	Markdown
Verify	thoughts/iterations/ITER-NNN.md (updated)	Markdown

Human Review Gate

Phase 4 (APPLY) requires explicit approval:

1. Review proposals in thoughts/proposals/
2. Confirm which changes to apply
3. Backups saved to thoughts/backups/

This ensures human oversight on all prompt changes.

Integration with dev-flow

dev-flow Component	Usage
dev_ledger	Track iteration tasks
dev_reasoning	Record iteration decisions
braintrust_analyze.py	Session data source (primary)
Local JSONL parsing	Session data source (fallback)

Skill Development Guide

See references/skill-development.md for:

• Creating new skills
• SKILL.md frontmatter best practices
• Skill quality checklist
• Plugin development guidelines

Local Mode

See references/local-mode.md for:

• Using without Braintrust
• Local data sources
• Troubleshooting common issues

References

• references/compound-learnings.md - Detailed process for transforming learnings into artifacts
• references/skill-template.md - SKILL.md template

Quick Reference

# Weekly check (full workflow)
/meta-iterate

# Find new skill opportunities
/meta-iterate discover

# Transform learnings into permanent artifacts
/meta-iterate compound

# After noticing issues
/meta-iterate evaluate --recent 5
/meta-iterate diagnose

# Apply improvements
/meta-iterate propose
# Review proposals...
/meta-iterate apply

# Verify after 5+ new sessions
/meta-iterate verify