Karpathy validated our entire thesis — here's the CLAUDE.md I use to run my trading business

On April 13, Andrej Karpathy pushed andrej-karpathy-skills to GitHub. By the end of the day, it had 5,828 stars. The repo is not a breakthrough model. It's not a new algorithm. It's a CLAUDE.md file and a handful of skills.

I'd been running a version of the same pattern for 6 months across six companies — an algo-trading service, a Claude Code OS product, a prompt-training platform, two daily newsletters, and a skills marketplace. The viral wave validated what I'd already built my business around: the instruction file IS the product.

Here's what's in mine. And a sanitized version you can download and adapt.

The insight behind CLAUDE.md

Every time you open Claude Code in a project, it reads CLAUDE.md at the repo root. Whatever's in that file is loaded into context before your first message. This is the cheapest, highest-leverage piece of context engineering available.

What most people miss: it's not about documentation. It's about encoding your operating principles as enforceable rules. Karpathy's file works because it tells Claude how to behave, not what the code does. The code can document itself. The rules cannot.

Anatomy of my trading business CLAUDE.md

Mine has four sections. Total length: about 350 lines. Loaded into context every session.

1. Critical Laws — non-negotiable rules

These are rules I've learned by burning capital when they were violated. Each one started as "I wish I'd known that" and graduated to "never again."

### Law #1: Think Before Act (UNIVERSAL)
Before writing ANY code, making ANY edit, or taking ANY action:
1. CLARIFY — What exactly is being asked? Re-read word by word.
2. MAP — Where does this belong? Which file, which section?
3. CHECK — What's the current state? Does this already exist?
4. PLAN — What files change? What needs verification after?
5. BUILD — Now write the code.
6. VERIFY — Re-read my exact words. Confirm the work matches.

This law alone changed the error rate on automated tasks by roughly an order of magnitude. Before: Claude would sometimes "fix" a bug by rewriting three unrelated files. After: Claude verifies the ask, maps it, and touches only what needs touching.

I have 26 laws like this. A few highlights:

• Law #2: Challenge & Verify — if I say something that can be data-checked, check it first. Don't blindly execute.
• Law #4: Recursive Quality — every error discovered becomes a permanent automated check, not a memory note.
• Law #10: Wrong Model Refusal — if the task requires higher reasoning than the current model tier, refuse and suggest the upgrade. Don't try and fail halfway.
• Law #21: Revenue-First Gate — until revenue > $0, every session must include at least one revenue-advancing action.
• Law #25: Visual QA Deploy Gate — no website, dashboard, or email ships without passing automated visual QA. Validators checking data structure are NOT sufficient. The gate must verify what the CUSTOMER sees.

2. Project-Specific Rules

Next: the invariants that matter for this repo. For a trading business:

• Ticker Segregation Law — every dashboard shows ONLY tickers from its designated universe. No leaking. Enforced at code level in generators, not just checklists.
• Shadow-First Process — ALL algo changes run in a parallel paper-trading shadow for 10+ days before they touch production.
• Template + Generator Architecture — never manually edit output HTML files. Always edit the generator or template.

These would be invisible to Claude without CLAUDE.md. It would happily hand-edit the rendered HTML. The rule says: don't.

3. Common Commands + Data Sources

Not a full reference — just the 10 commands and 5 data files Claude hits most often. This cuts the "which script was that?" exploration time to zero.

## Common Commands
- Run command center: `python3 code/generate_command_center.py`
- Validate: `python3 code/validate_ep_dashboard.py`
- Run shadow pipeline: `python3 code/run_shadow.py`
- Push external: `python3 code/generate_ept_external.py --push`

## Data Sources
- `code/outputs/signals.json` — main pipeline output
- `brain/adj_config.json` — data-driven ADJ parameters
- `brain/dashboard_spec.json` — canonical dashboard spec

4. EP Preferences

Personal defaults. Not rules — preferences. Terse responses, always include reasoning, never assume "test data," always create a backup before writing. Claude honors these once they're written.

What Karpathy gets right

The Karpathy repo's core insight, translated for anyone running a business:

A good CLAUDE.md is not documentation. It is a contract between you and the AI about how work gets done in this codebase.

Every time I've watched someone try to use Claude Code effectively without a real CLAUDE.md, I see the same pattern: Claude produces technically-correct code that violates three unstated rules. The engineer gets frustrated, adds the rule to their prompt, Claude obeys for that session, and the rule is lost next session.

CLAUDE.md is the answer. But most CLAUDE.md files I see in the wild are a paragraph about what the app does. That's useless. The valuable content is the enforceable instructions — what to do, what not to do, in what order.

Download the sanitized version

If this is interesting, you'd probably like Brainfile

The sanitized download above is the skeleton. What I actually sell — Brainfile — is the fully operationalized version: your CLAUDE.md, plus the agents that execute it, plus the integrations that connect it to Stripe / Resend / your domain / your CI. A managed Claude Code operating system for a business, not just an instruction file.

Karpathy just proved 5,828 developers want the pattern. Brainfile is for the ones who want it ready to run.

See Brainfile →