Print Money Factory

An npm package for Claude Code that provides a complete AI-driven trading strategy development pipeline via slash commands (/brrr:*). Describe a trading idea, and the system iteratively develops, backtests, optimizes, and exports it — from hypothesis to ready-to-trade PineScript and Python code.

How It Works

/brrr:new-milestone  →  Define your strategy idea and scope
/brrr:discuss        →  Fix all decisions: entry/exit, stops, sizing, params
/brrr:research       →  Find implementations, pitfalls, lookahead traps
/brrr:plan           →  Design parameter space and optimization method
/brrr:execute        →  AI writes & runs backtests, analyzes, adjusts, repeats
/brrr:verify         →  Interactive HTML report, approve or debug

The core loop: idea → backtest → AI analysis → parameter adjustment → repeat until targets are hit or the strategy is diagnosed as unviable.

Installation

npx @print-money-factory/cli install

Requirements: Python 3.10+ (the only prerequisite — everything else is installed automatically)

What it does:

• Copies 9 slash commands to ~/.claude/commands/brrr/
• Creates a Python venv at ~/.pmf/venv/ with all dependencies
• Installs workflows, templates, and reference modules to ~/.pmf/

After install, open Claude Code and type /brrr:new-milestone to start.

Commands

| Command | What it does | |---------|-------------| | /brrr:new-milestone | Create a new strategy milestone with guided scoping | | /brrr:discuss | Fix strategy decisions through conversation (entry/exit, stops, sizing, commissions, parameter ranges) | | /brrr:research | Find known implementations, academic work, and lookahead traps for your strategy type | | /brrr:plan | Design parameter space, select optimization method (grid/random/walk-forward/bayesian), set evaluation criteria | | /brrr:execute | Run the AI backtest loop: load data → run backtest → compute metrics → AI analyzes → adjust params → repeat | | /brrr:verify | Generate interactive HTML report with equity curve, drawdown, regime breakdown, benchmark comparison | | /brrr:status | ASCII tree showing milestone progress, all phases, next step | | /brrr:doctor | Diagnose installation health (Python, venv, dependencies, file integrity) | | /brrr:update | Update to the latest version |

The Pipeline

1. New Milestone (/brrr:new-milestone)

Define your trading idea and scope. The system guides you through:

• Strategy idea (e.g., "SMA crossover on BTC/USDT daily")
• Scope selection (strategy, backtest, tuning, risk management, exports)
• Success criteria (target Sharpe, max drawdown, min trades)
• Asset and data source selection

2. Discuss (/brrr:discuss)

Fix all strategy decisions before any code runs:

• Entry/exit logic, stop-loss, take-profit
• Position sizing and commission assumptions
• Parameter ranges for optimization
• --auto flag for reasonable defaults with minimal questions

Debug mode: When coming back from a failed verify, discuss starts from the AI diagnosis — not from scratch. It reads all prior failed approaches and avoids suggesting what already didn't work.

3. Research (/brrr:research)

Optional phase that investigates your strategy type:

• Known implementations and academic references
• Lookahead bias traps specific to your strategy
• --deep flag for extended web search

4. Plan (/brrr:plan)

Design the optimization approach:

• Define free parameters with ranges and step sizes
• Auto-selects optimization method:
  • Grid search for < 100 combinations
  • Random search for 100-500 combinations
  • Bayesian (Optuna TPE/CMA-ES) for 500+ combinations
  • Walk-forward as override option
• Set evaluation criteria, data period, train/test split
• Parameter budget enforcement to prevent overfitting

5. Execute (/brrr:execute)

The core AI backtest loop — fully autonomous:

1. Claude writes a Python backtest script from scratch based on your plan
2. Runs it against real market data (ccxt for crypto, yfinance for stocks)
3. Computes 9 metrics: Sharpe, Sortino, Calmar, Max DD, Win Rate, Profit Factor, Expectancy, Trade Count, Net PnL
4. AI analyzes results and equity curve PNG
5. Adjusts parameters and repeats

Stop conditions:

• MINT — All targets hit
• PLATEAU — 3 iterations without >5% improvement
• REKT — No edge found at any parameter combination
• NO DATA — Data loading failed

Bayesian optimization: When selected in the plan phase, uses Optuna's Ask-and-Tell API with TPE or CMA-ES sampler. Each iteration shows [WARMUP] or [GUIDED] mode. Studies persist to SQLite for --resume.

Per-iteration artifacts saved: params JSON, metrics JSON, equity curve PNG, verdict JSON.

6. Verify (/brrr:verify)

Generates a standalone interactive HTML report (Plotly, no server needed) with:

| Section | What it shows | |---------|--------------| | Metrics Summary | All 9 metrics vs targets with color coding | | Equity Curve | Strategy vs buy-and-hold with zoom | | Drawdown Chart | With max drawdown horizontal line | | Iteration Table | All iterations with Sharpe evolution | | Parameter Heatmap | If grid search was used | | Trade List | Per-trade P&L with green/red coloring | | Regime Breakdown | Bull/bear/sideways performance | | Benchmark | Alpha, beta vs buy-and-hold |

Three paths:

• --approved — Closes the milestone and generates a complete export package (8 files)
• --debug — AI diagnoses what went wrong, writes a structured diagnosis JSON with "do NOT retry" entries, opens a new phase cycle
• Interactive — AI presents assessment, you choose

Export Package (on --approved)

Generated in .pmf/output/:

| File | Purpose | |------|---------| | pinescript_v5_strategy.pine | TradingView strategy for backtesting | | pinescript_v5_indicator.pine | TradingView indicator with alerts for live trading | | trading-rules.md | Plain English entry/exit/sizing rules | | performance-report.md | Portable metrics summary | | backtest_final.py | Reproducible Python script, runs standalone | | live-checklist.md | Step-by-step guide before real money | | bot-building-guide.md | Platform-specific deployment instructions (crypto/stocks/forex) | | report_vN.html | Interactive HTML report |

Data Sources

| Source | Assets | Timeframes | API Key | |--------|--------|------------|---------| | ccxt | Crypto (100+ exchanges) | All | No | | yfinance | Stocks, ETFs, Forex (daily) | Daily | No | | CSV | Any asset | Any | No | | polygon.io | Stocks (intraday) | Minute+ | Yes |

Data is validated for gaps and NaN before every backtest run. Local caching prevents re-downloading.

Optimization Methods

| Method | When | How | |--------|------|-----| | Grid Search | < 100 combinations | Systematic sweep through all parameter combinations | | Random Search | 100-500 combinations | Random sampling guided by AI analysis | | Bayesian (Optuna) | 500+ combinations | TPE sampler (mixed params) or CMA-ES (continuous only), SQLite persistence for --resume | | Walk-Forward | Time-series robustness | Rolling train/test windows |

Debug Cycles

When a strategy doesn't meet targets, the debug cycle carries forward knowledge:

1. /brrr:verify --debug writes a structured phase_N_diagnosis.json with:

   • Failed parameter regions
   • Explicit "do NOT retry" entries
   • AI-generated analysis of what went wrong

2. /brrr:discuss in debug mode:

   • Reads ALL prior diagnosis files
   • Presents a "Prior Debug Cycles" failure summary table
   • Enforces do_not_retry constraints — the AI cannot suggest parameters that already failed
   • Starts from the diagnosis, not from scratch

3. Hypothesis drift protection detects when changes exceed the original strategy scope and offers to open a new milestone instead.

Architecture

~/.claude/commands/brrr/    # 9 thin slash commands (markdown)
~/.pmf/
  workflows/                # 8 behavioral workflows (markdown prompts)
  references/               # Python modules (metrics, backtest engine, data sources, optuna bridge, report generator)
  templates/                # HTML report, PineScript, STATE.md, STRATEGY.md
  venv/                     # Isolated Python environment
  .version                  # Installed version info

Commands are thin markdown files that @-reference behavioral workflows. The workflows describe what Claude does step-by-step. Python modules are fixed reference code (not regenerated per strategy) with unit tests.

Metrics

All 9 metrics computed per iteration with known-answer TDD tests:

| Metric | What it measures | |--------|-----------------| | Sharpe Ratio | Risk-adjusted return (annualized) | | Sortino Ratio | Downside risk-adjusted return | | Calmar Ratio | Return vs max drawdown | | Max Drawdown | Largest peak-to-trough decline | | Win Rate | Percentage of winning trades | | Profit Factor | Gross profit / gross loss | | Expectancy | Average profit per trade | | Trade Count | Total number of trades | | Net PnL | Total profit/loss |

Updating

/brrr:update

Or manually:

npx @print-money-factory/cli@latest install

Every /brrr:* command silently checks for updates once per 24 hours.

Diagnostics

/brrr:doctor

Checks: Python version (>=3.10), venv health, 10 library imports, command files, workflow files, reference files. Reports pass/fail per check with fix suggestions.

Requirements

• Python 3.10+ (tested on 3.10 through 3.14)
• Claude Code (the CLI tool from Anthropic)
• Node.js 18+ (for the install script)

All Python dependencies are installed automatically in an isolated venv: pandas, numpy, matplotlib, plotly, optuna, ccxt, yfinance, scipy, ta, jinja2, tabulate, requests, pyyaml, pytest

Design Principles

• One milestone at a time — Keeps focus, prevents strategy confusion
• Claude writes the backtest engine — No framework constraints, maximum flexibility per strategy
• Behavioral workflows — Claude follows step-by-step markdown prompts, not code templates
• Per-iteration artifacts — Every iteration saves params, metrics, equity PNG, verdict JSON
• Standalone outputs — HTML reports work without a server, PineScript pastes directly into TradingView

License

MIT