karpathy-guidelines

Karpathy Guidelines

Four principles that prevent the most common AI-assisted coding failure modes. Load this skill when reviewing implementation plans or code output.

Rules

Rule 1 — Think Before Coding

Spend time understanding the problem before writing a single line. Re-read the issue, the acceptance criteria, and any relevant ADRs.

Bad:

// Immediately add a retry wrapper because the API call might fail
async function fetchWithRetry(url) { … }

Good:

// 1. Read the issue: the API is internal, on the same VPC. Retries not needed.
// 2. Check ADR-0005: we use circuit-breakers at the gateway, not per-caller.
// Conclusion: a plain fetch() is correct here.
async function fetchConfig(url) { return fetch(url).then(r => r.json()); }

---

Rule 2 — Simplicity First

Do not add abstraction until you have two concrete use cases that share it. The cost of wrong abstraction is higher than the cost of duplication.

Bad:

// Premature factory for a single use case
class DataFetcherFactory {
  static create(type: 'user' | 'post'): DataFetcher { … }
}

Good:

// Concrete function for the one case that exists today
async function fetchUser(id: string): Promise<User> { … }
// If fetchPost() is needed later and shares logic → extract then, not now.

---

Rule 3 — Surgical Changes

Change only what the task requires. Do not refactor adjacent code, rename variables "while you're here", or fix unrelated linting warnings in the same commit.

Bad:

Task: "Add timeout to fetchData()." Agent refactors all callers, introduces RetryConfig class, renames three variables for clarity, fixes unrelated import order.

Good:

Task: "Add timeout to fetchData()." Agent adds one parameter with a default, threads it to the fetch call, updates the one test that exercises fetchData. Diff: 4 lines.

---

Rule 4 — Goal-Driven Execution

Keep the original task visible. Before each action, ask: "Does this directly advance the stated goal?" Stop when the goal is met — do not add polish, documentation, or "nice to have" features beyond scope.

Bad:

// Task complete, but agent adds:
// - JSDoc on every function (not requested)
// - A README section for the new parameter (not requested)
// - A CHANGELOG entry (not requested)

Good:

// Task complete. Verification run. No out-of-scope additions.
// Noted potential improvement → opened separate issue #47.

---

When to Apply These Rules

• Before starting any implementation: apply Rule 1 and Rule 4 (understand goal, confirm scope).
• During implementation: apply Rule 3 (surgical).
• During review of your own output: apply Rule 2 (simplicity check).

See REFERENCE.md for anti-pattern catalog and concrete scenarios.