Karpathy open-sourced the methodology. We turned it into 12 ready-to-run program.md templates that optimize your cold email, landing pages, and pricing — on autopilot, at near-zero cost with local models.
Already sold? Skip to the full playbook — $97 →
The open-source repo needs a GPU, PyTorch, and Python. We translate the exact same optimization loop for marketing and sales operators — no terminal heroics required.
You can't point a GPU training loop at a landing-page headline or a cold-email subject line. This playbook bridges that gap with business-ready instructions.
Other guides explain what loops are. We hand you the exact templates: copy, fill in your metrics, and run on Claude Code or Cursor.
Every loop is framed in business terms: define what can change, pick the success scorecard, set the safety limits, keep the winner, repeat.
Everything you need to launch and run compounding AI loops — without writing a single script.
Ready-to-run instruction sets for local or cloud agents — covering cold email, headline testing, pricing realization, and onboarding activation.
Route your loops to local models (Llama 3, Mistral, Gemma) via Ollama and MLX, and run unlimited experiments at zero marginal API cost.
The exact ways agent loops waste money — changing two variables at once, undersized samples, optimizing vanity metrics — and how to avoid each.
This is what an instruction set looks like. Switch tabs to see how the same loop adapts across domains.
On June 15, 2026, Anthropic moves programmatic agent usage off flat subscriptions and onto metered API billing. If you run loops nightly, here's the real math:
100 iterations a night at Opus 4.8 rates (~$5 / $25 per million tokens), metered per token, every night.
Run the identical loop on local models (Llama 3, Mistral, Gemma) on your own machine or Apple Silicon.
No setup marathon. Three moves and the agent is running.
Pick the domain with the fastest feedback — usually cold email or landing pages — and paste the matching program.md into a file.
Replace the bracketed placeholders with your current metrics, your offer, and your sending tool. Specific inputs, sharper output.
Tell Claude Code or Cursor to read program.md and start the loop. Come back in 24–48 hours and feed it the results.
Try the templates for two weeks. If you don't ship at least one optimization you'd have missed, reply to your receipt and we refund 100% — no questions, no conditions.
One-time payment. Lifetime updates. Pick the bundle that matches how much hand-holding you want.
Karpathy proved it with code. Now prove it with your business.
Get the Playbook — $97 →