← ClaudeAtlas

karpathy-guidelineslisted

Behavioral coding-agent guardrails adapted from the Karpathy-inspired Claude/Cursor skill archive. Use when writing, reviewing, refactoring, debugging, or planning code changes where Codex should surface assumptions, avoid overengineering, keep diffs surgical, preserve unrelated user work, and define verifiable success criteria before declaring the task done.
andrewr303/claude-codex-plugin-lab · ★ 0 · Code & Development · score 60
Install: claude install-skill andrewr303/claude-codex-plugin-lab
# Karpathy Guidelines Use this skill as a lightweight behavioral overlay for coding work. It does not replace project instructions, PRDs, schemas, security rules, or local harness docs. In HypeStake, repo-specific AGENTS/CLAUDE/wiki guidance always wins when there is a conflict. ## Operating Loop 1. Name the target in implementation terms. 2. State assumptions that affect scope, data access, permissions, or user-visible behavior. 3. If the request has materially different interpretations, surface the tradeoff before editing. Ask only when a reasonable assumption would be risky, destructive, or product-defining. 4. Choose the smallest source-backed change that solves the target. 5. Define success criteria and verification before calling the task done. ## Four Guardrails ### 1. Think Before Coding - Do not silently choose between plausible meanings when the difference matters. - Make uncertainty visible early. - Push back when the user's requested path is broader, riskier, or more complex than needed. - Read the relevant source files before inventing behavior. ### 2. Simplicity First - Do not add features, switches, abstractions, or configuration that the task did not require. - Prefer existing local helpers and patterns. - Avoid one-use abstractions unless they remove real complexity. - If a solution feels large for the ask, pause and simplify. ### 3. Surgical Changes - Touch only files and lines needed for the requested behavior. - Match the surrounding style even w