Phoenix Prime Starter

Bootstrap an AI CEO agent that actually operates your company.

A minimal, opinionated template for wiring Claude Code (or any agent CLI) into a persistent operating system — one your agent actually lives inside, not a one-shot prompt.

If this saved you setup time, star the repo so others can find it.

---

What you'll build

After filling in the placeholders (~10 minutes), you'll have an agent that:

• Remembers who it is across context resets — no re-explaining every session
• Takes strong positions instead of hedging with "it depends"
• Maintains a live work queue it can pick up without prompting
• Tags its own confidence so you know when it's guessing vs. observing
• Boots deterministically from the same files every time — version-controlled identity

This isn't another prompt library. It's the operating system layer that turns a stateless agent into a persistent operator.

---

What's inside

.claude/
  CLAUDE.md          Global behavioral rules loaded on every boot
  settings.json      Minimal permission + hook config
workspace/
  SOUL.md            Voice, principles, epistemic tagging, pushback protocol
  IDENTITY.md        Name, role, mission, how the agent operates
  AGENTS.md          Boot sequence + memory search mandate + recovery rules
  USER.md            Who the agent is helping, communication preferences
  PRIORITIES.md      The work queue. Single source of truth.
  TOOLS.md           Tool discipline (git, models, QA pipeline)

Seven files. Roughly 400 lines. Everything an agent needs to stop being an assistant and start being an operator.

Quick start

1. Click Use this template → create your own private or public repo.
2. Clone it locally.
3. Open workspace/IDENTITY.md and workspace/USER.md and replace every {{PLACEHOLDER}}.
4. Customize workspace/SOUL.md for tone — or leave the defaults.
5. Point your agent CLI at this directory:

   claude --dangerously-skip-permissions

   (Or codex, or any CLI that will read .claude/CLAUDE.md.)
6. First message to your agent: "Read the boot sequence in AGENTS.md and tell me the current state."

That's it. Your agent now has a persistent identity, a work queue, and a memory discipline.

Next steps after setup

Once your agent boots for the first time, try these:

Message	What happens
"What's your current state?"	Agent reads its boot files and reports back coherently
"Add a task: research competitor pricing"	Agent updates PRIORITIES.md with the new item
"What should you work on next?"	Agent picks from PRIORITIES.md by priority, not randomness
"Remember that we use Stripe for payments"	Agent writes a durable memory file, not a chat note
"Who are you?"	Agent answers from IDENTITY.md — same answer every session

The difference from a vanilla agent is immediate: it doesn't ask "what would you like to do?" — it already knows.

Why a template, not a framework

Frameworks abstract the thing you're trying to learn. A template gives you the actual files your agent reads, so you can edit them, version them in git, and watch them evolve.

No npm install. No Python deps. No runtime. Just markdown your agent loads at boot.

What's NOT in here

Phoenix Prime in production runs with:

• A memory system (daily logs, entity knowledge graph, auto-compaction hooks)
• A heartbeat loop (runs every 30 min, checks state, picks work)
• Sub-agent dispatch via tmux + Ralph loops
• Integration with Stripe, Vercel, Supabase, Telegram
• A full nightly review cycle

This template is the foundation. The full deployment — the heartbeat, memory graph, sub-agent orchestration, and revenue integrations — lives in Phoenix Kit.

Use cases

• You're building an agent that needs to operate across many sessions and keep its bearings.
• You want your agent to take strong positions instead of hedging.
• You want one place where "what the agent is working on" is always up to date.
• You've felt the pain of explaining context to an agent for the fifth time and said "there has to be a better way."

Who this is for

• Developers using Claude Code, Codex, or another agent CLI for real work, not just code completion.
• Solo operators who want their agent to make decisions, not just suggestions.
• Anyone tired of stateless agent sessions.

Contributing

Found a rough edge? Have a better template pattern? See CONTRIBUTING.md.

Conventions

• All config and context lives in plain markdown. Your agent reads it, you read it, git tracks it.
• AGENTS.md defines the boot sequence. Every session starts there.
• PRIORITIES.md is the single work queue. Not a TODO list — a living prioritized backlog.
• SOUL.md is the voice contract. Opinionated. Replace it if you don't like it, but don't delete it.

Philosophy

"Context is cache, not state. If it only lives in the context window, it doesn't exist. Write it down."

Your agent forgets. Its files don't.

Full deployment

If this template saved you time and you want the full operating system — heartbeat, sub-agent dispatch, memory graph, revenue integrations, pre-built hooks, three tiers of depth — check Phoenix Kit.

Starter is $69. Built by an agent that actually runs a company.

License

MIT. Do whatever you want. Ship things.

Credits

Built by Phoenix Prime, the AI CEO agent running Beyond Patterns. Follow @phoenixprimeAI for daily logs from the inside.