The ARENA Playbook Blueprint
An AI beat ChatGPT, Claude, Gemini and Grok in a real-money trading contest: +22% in 17 days. I rebuilt its playbook as a 4-agent team. Here's how it works and a free prompt to run it in 10 minutes.
How it works: Market data → 4 agents (scout, analyst, risk, judge) → one trade plan with a stop → you decide
READ FIRST: It's a research playbook, not a promise. It never places a trade. Use small risk on anything you try.
Free full guide (step by step, with 12 copy-paste prompts, daily routine and weekly review): milli-ai.com/free/arena
Get the full free guide →How it works
- SIGNAL checks Bitcoin, Ethereum and Solana every hour (free data). It only looks for the two setups above. No setup? No AI call, no cost.
- JUDGE is Claude. It scores the setup and sets the stop and target. It's told to skip when in doubt.
- RISK is plain code. The AI can't argue with it. Score under 70, bad stop, small reward, or 3 trades already today? Blocked. Every trade is sized so a loss costs just 1%.
- JOURNAL saves every idea and texts the ticket to your phone.
After a month, the log shows if the high-scoring trades really did better. That's how you tune it.
Educational only. Not financial advice. Crypto moves fast and you can lose money.
Build the free starter version with Claude Code (20 minutes)
Want it running on your computer instead of in a chat? Paste one prompt and Claude Code builds a starter ARENA crypto desk: SIGNAL, JUDGE, RISK and JOURNAL agents on BTC, ETH and SOL. It runs on pretend money or in read-only mode. It never places a trade or bet.
You need: Claude Code (needs a paid Claude plan), Python 3, and an Anthropic API key from console.anthropic.com (pay as you go, usually cents per run).
1. Install Claude Code (Mac/Linux terminal):
curl -fsSL https://claude.ai/install.sh | bash
Windows PowerShell: irm https://claude.ai/install.ps1 | iex
2. Make a folder and open Claude Code in it:
mkdir arena && cd arena
claude
3. Paste this build prompt:
Build me a beginner-friendly starter bot called arena in Python 3: a starter ARENA crypto desk: SIGNAL, JUDGE, RISK and JOURNAL agents on BTC, ETH and SOL.
WHAT IT DOES
1. Pull daily candles for BTC, ETH and SOL from a free public source (for example the Coinbase Exchange public candles endpoint). If unsure of an endpoint, leave a TODO.
2. SIGNAL (Python): trend and 20-day breakout. JUDGE (Claude): builds a plan with entry, stop, target. RISK (Python): size at 1% of a pretend $10,000 account, reject if reward under 2x risk, no leverage.
3. Pretend-money only. Log every plan and result to journal.jsonl.
RULES FOR THE BUILD
4. Use the official Anthropic Python SDK. Put the model name in config.json. Read ANTHROPIC_API_KEY from .env with python-dotenv. Create .env.example with placeholders and add .env, data/ and logs/ to .gitignore. Never hardcode keys.
5. It never places real trades or bets. No broker or wallet code, no private keys.
6. Put every setting (limits, tickers, thresholds) in config.json so I can change them without touching code.
7. Add a --demo flag that runs the whole thing on made-up data with no internet, so I can test it first.
8. Handle errors and rate limits with a short wait and retry. If you are not sure an API endpoint or field exists, leave a clear TODO instead of guessing.
9. Write a README.md for a total beginner: install, add the key, run --demo, run for real, read the logs.
10. Keep it simple and well commented, under about 400 lines. When done, explain what you built and list anything you could not check.
4. Install and test:
pip install -r requirements.txt
cp .env.example .env
python src/main.py --demo
Add your key to .env with a text editor, then run it for real (the README shows the exact command).
Keep your key safe: never paste API keys into a chat, a screenshot or GitHub. Keep
.envin.gitignore.
Starter vs finished
| Starter (free, this prompt) | Finished (AI Desk) | |
|---|---|---|
| Works on your computer | ✅ | ✅ |
| Runs on a schedule with phone alerts | ❌ | ✅ |
| Tested risk rules | ❌ | ✅ |
| Weekly journal review | ❌ | ✅ |
| Multi-model judge | ❌ | ✅ |
| Tested by us, with setup support | ❌ | ✅ |
Troubleshooting
| Problem | Fix |
|---|---|
| The AI makes up prices | Paste the real prices yourself and say "use only these numbers". |
| The plan has no stop | Reply: "Every plan needs a stop. Redo it." |
| All 4 agents agree every time | Ask the risk agent to argue against the trade first. |
Want the finished version?
The prompt above builds a starter. The finished, tested version (one-command install, plain setup steps, and help if you get stuck) is part of the Milli.ai AI Desk.
- Free full guide: milli-ai.com/free/arena
- Finished bot: milli-ai.com
Educational only. Not financial advice. It never places trades for you. You make every decision.