andrej-karpathy-skill
Apply Andrej Karpathy-inspired coding-agent guidelines in Codex. Use when writing, reviewing, debugging, or refactoring code to surface assumptions, avoid overengineering, keep edits surgical, and define verifiable success criteria.
Works with
--- name: andrej-karpathy-skill description: Apply Andrej Karpathy-inspired coding-agent guidelines in Codex. Use when writing, reviewing, debugging, or refactoring code to surface assumptions, avoid overengineering, keep edits surgical, and define verifiable success criteria. license: MIT --- # Andrej Karpathy Skill Use this skill as a Codex-native version of the Karpathy coding-agent guidelines. The goal is not to add a new framework. The goal is to make Codex behave more carefully on real code: clarify before guessing, prefer simple implementations, avoid unrelated edits, and verify against a concrete goal. ## The Four Checks ### 1. Think Before Coding Before editing, make the task explicit. - State the interpretation you are using. - Surface assumptions that affect the implementation. - Name meaningful tradeoffs when more than one path is reasonable. - Ask a concise clarifying question when guessing would create real risk. - If the task is obvious and low-risk, state the assumption briefly and proceed. Codex should not silently pick a risky interpretation and run with it. ### 2. Keep It Simple Implement the smallest thing that satisfies the current request. - Do not add features the user did not ask for. - Do not add configurability before there is a real need. - Do not create abstractions for one caller. - Do not introduce new dependencies for logic the repo can already express simply. - If the first approach feels like architecture, look for the direct version first. Codex should solve today's problem, not design tomorrow's system by accident. ### 3. Make Surgical Changes Keep the diff tied to the request. - Touch only the files needed for the task. - Match the local style even when another style is personally preferable. - Do not reformat, rename, or reorganize adjacent code as a side effect. - Clean up imports, variables, or helpers made unused by your own change. - Mention unrelated dead code or design problems separately instead of fixing them inside the patch. Codex should leave the surrounding code recognizable. ### 4. Define The Goal And Verify It Turn the request into a checkable outcome before calling it done. - Bug fix: identify the failing case and the expected behavior. - Feature: identify the behavior the user should be able to observe. - Refactor: identify the behavior that must remain unchanged. - Review: identify concrete risks, missing tests, and regressions. Use the narrowest meaningful verification available. If you do not run a check, say so plainly and explain why. ## Codex Response Pattern For non-trivial coding work, keep the user oriented with: ```text Assumption: Changed: Verified: Remaining risk: ``` Use this shape lightly. Do not add ceremony to obvious one-line edits. ## Pushback Push back gently when the request or your first design would cause avoidable scope growth: - a broad rewrite for a narrow bug, - a new abstraction with one use case, - a formatting sweep mixed into behavior changes, - a public API expansion that is not required, - a verification plan too weak for a risky change. When pushing back, offer the smaller path that still satisfies the user's goal.
More Debugging skills
diagnosing-bugs
mattpocock/skills
Diagnosis loop for hard bugs and performance regressions. Use when the user says "diagnose"/"debug this", or reports something broken/throwing/failing/slow.
explore-code
lllllllama/rigorpilot-skills
Rigor Improve implementation leaf skill for auditable candidate implementation in deep learning research repositories. Use when the researcher explicitly authorizes exploratory work on an isolated branch or worktree to transplant modules, adapt a backbone, add LoRA or adapter layers, replace a head, or stitch together meaningful low-risk migration ideas with rollback-aware records in `explore_outputs/`. Do not use for end-to-end exploration orchestration on top of `current_research`, trusted baseline reproduction, conservative debugging, environment setup, verified contribution claims, or default repository analysis.
safe-debug
lllllllama/rigorpilot-skills
Rigor Debug / Rigor Audit skill for deep learning research work. Use when the user pastes a traceback, terminal error, CUDA OOM, checkpoint load failure, shape mismatch, NaN loss symptom, or training failure and wants conservative diagnosis before any patching, with debug fixes clearly separated from research contributions. Do not use for broad refactoring, speculative adaptation, automatic exploratory patching, or general repository familiarization.

