| name | karpathy-guidelines |
| description | Apply four anti-error coding principles plus this repository's stricter financial-evidence, safety, and single-session rules. |
| license | MIT |
| metadata | {"origin":"multica-ai/andrej-karpathy-skills at 2c606141936f1eeef17fa3043a72095b4765b9c2"} |
Karpathy Guidelines
Apply these principles before the repository-specific overlay. They reduce
wrong assumptions and unnecessary code; they do not replace risk controls or
verification.
Four Principles
- Think before coding. State uncertain assumptions, expose conflicting
interpretations, and verify unfamiliar behavior before choosing a path.
- Prefer the simplest complete solution. Add no speculative abstraction or
configuration. Complexity must remove demonstrated risk or duplication.
- Make surgical changes. Preserve unrelated work, match established local
contracts, and remove only orphans created by the current change.
- Work to verifiable outcomes. Define the behavior and evidence that prove
success, then loop until the relevant checks pass.
Trading Overlay
- Never infer profitability, financial edge, risk containment, data quality, or
production readiness. Require reproducible artifacts and independent gates.
- Keep signed execution testnet-only and fail closed on uncertain ownership,
stale market data, incomplete reconciliation, exhausted API budget, or
invalid model evidence.
- Preserve CLI/Windows functional parity through generated contracts.
- Use focused checks while iterating and full verification at promotion or
release boundaries.
- Never expose credentials or use another agent/session.
AGENTS.md is the
authoritative stricter policy.