agent-seed

Self-evolving AI assistant starter kit. Markdown-only. Works with any LLM agent.

Plant the seed. It grows with you.

---

What is this?

A collection of markdown files that teach AI agents who you are — your voice, your preferences, your work context, your priorities.

It works with Claude Code, Codex, Devin, Cursor, Windsurf, and anything else that reads markdown.

No API keys. No runtime. No vendor lock-in. Just .md files.

What it does

• Gives any AI your voice — communication style, tone, patterns
• Tracks your daily priorities — todos, drafts, notes
• Self-improves — learns from each session, proposes new agents via PRs
• Stays small — enforced size limits prevent markdown bloat
• Grows with you — /setup generates agents based on YOUR workflow, /learn evolves them over time

What it doesn't do

• Doesn't require API keys or expose secrets
• Doesn't lock you into any tool or vendor
• Doesn't run any code — it's just context files

Quick Start

1. Use this template

Click "Use this template" on GitHub, or:

git clone https://github.com/jcurbelo/agent-seed.git my-agent
cd my-agent
rm -rf .git && git init

2. Run setup

Open the repo in your preferred AI coding tool and run:

/setup

The setup wizard will ask about your name, role, communication style, and tools — then generate personalized agents, voice profiles, and context files.

3. Start using it

/standup    # Morning briefing — scan todos, plan the day
/review     # Quick check-in — what needs attention?
/eod        # End of day — capture learnings, prep tomorrow
/learn      # Self-improvement — prune, evolve, grow new agents

How It Works

The Index Pattern

Every folder has an index.md that links to detail pages — like imports in code:

context/
├── index.md              # Master index → links to voice/, work/
├── voice/
│   ├── index.md          # Voice overview → links to patterns.md
│   └── patterns.md       # Detailed examples
└── work/
    ├── index.md          # Role overview → links to tools.md
    └── tools.md          # Services and repos you use

Index files are always loaded first. Detail pages load on demand. Size limits keep everything fast.

Self-Evolution

This is the core idea. Three layers:

1. Reflection agent — after meaningful sessions, captures learnings to memory/
2. /learn skill — periodic self-improvement: prunes stale entries, identifies gaps, creates new agents via PR
3. /setup re-run — add new integrations or agents anytime

When the system detects a gap (e.g., you keep asking about Slack but there's no Slack agent), it creates one and files a PR.

Frontmatter

Every .md file has YAML frontmatter that defines its role:

# Agent definition
---
model: sonnet                    # Model tier: haiku | sonnet | opus
tools: [Read, Write, Edit]      # Allowed tools
permissionMode: default          # default | bypassPermissions
maxTurns: 10                     # Max agent turns
---

# Scoped rule
---
paths: ["todos/**"]              # When these paths are active
---

# Context/memory page
---
title: Voice Profile
description: Communication style and tone
parent: ../index.md              # Navigation link
max_lines: 250                   # Enforced by policy script
---

Markdown Policy

A script (scripts/check-markdown-policy.sh) enforces:

• Size limits per file type (index: 150 lines, pages: 250 lines, memory: 200 lines)
• Required frontmatter keys
• Required todo sections
• No orphaned pages (every page linked from an index)

Run it locally or add it to CI.

File Structure

agent-seed/
├── CLAUDE.md                     # Claude Code / LLM entrypoint
├── AGENTS.md                     # Devin / Codex entrypoint
├── .cursorrules                  # Cursor entrypoint
├── README.md                     # You are here
├── LICENSE                       # MIT
│
├── .claude/
│   ├── settings.json             # Safety hooks
│   ├── agents/                   # Agent definitions
│   │   ├── index.md              # Agent registry + routing
│   │   ├── researcher.md         # Web research
│   │   ├── coder.md              # Code writing
│   │   ├── reviewer.md           # Code review
│   │   ├── todo-manager.md       # Daily todos
│   │   └── reflection.md         # Self-improvement engine
│   ├── skills/                   # Slash commands
│   │   ├── setup/SKILL.md        # /setup wizard
│   │   ├── standup/SKILL.md      # Morning briefing
│   │   ├── eod/SKILL.md          # End of day
│   │   ├── review/SKILL.md       # Quick check-in
│   │   ├── learn/SKILL.md        # Self-improvement
│   │   └── update-todos/SKILL.md # Interactive todo review
│   └── rules/                    # Scoped rules (load by file path)
│       ├── todos.md
│       ├── context.md
│       └── drafts.md
│
├── context/                      # Who you are (populated by /setup)
│   ├── index.md
│   ├── voice/
│   │   ├── index.md
│   │   └── patterns.md
│   └── work/
│       ├── index.md
│       └── tools.md
│
├── memory/                       # What the AI has learned
│   ├── MEMORY.md                 # Master index (always loaded)
│   ├── learnings.md
│   ├── preferences.md
│   ├── patterns.md
│   └── improvements.md
│
├── todos/                        # Daily task tracking
│   └── _template.md
│
└── scripts/
    └── check-markdown-policy.sh  # Enforces structure + size limits

Compatibility

Tool	Entrypoint	Status
Claude Code	CLAUDE.md	Full support (agents, skills, rules, hooks)
Cursor	.cursorrules	Context + rules
Codex	AGENTS.md	Context + agents
Devin	AGENTS.md	Context + agents
Windsurf	CLAUDE.md	Context + rules
Any LLM	CLAUDE.md	Read as context

Philosophy

Your AI assistant should know you the way a good colleague does — your communication style, your priorities, your tools. This repo is that knowledge, version-controlled and self-evolving.

Use it with current AI coding agents. Don't expose private keys. Do whatever you want with it.

License

MIT — do whatever you want.