From karpathy-skills
Karpathy-inspired behavioral guidelines (authored by forrestchang/multica-ai, not by Andrej Karpathy) to reduce common LLM coding mistakes. Four principles — think before coding, simplicity first, surgical changes, goal-driven execution — applied during writing, reviewing, or refactoring code to avoid overcomplication, surface assumptions, and define verifiable success criteria. Use when starting a coding task and you want a short behavioral checklist, when reviewing code that feels overengineered, or when scoping a refactor to keep it surgical. Activation hints: "karpathy guidelines", "llm coding pitfalls", "behavioral guidelines for coding", "simplicity first", "surgical change", "think before coding", "goal-driven execution".
How this skill is triggered — by the user, by Claude, or both
Slash command
/karpathy-skills:karpathy-skillsThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
> **Learnings (Maintainer):** bekannte Probleme/Erkenntnisse zu diesem Skill in `docs/skill-learnings.md` → Sektion „karpathy-skills".
Learnings (Maintainer): bekannte Probleme/Erkenntnisse zu diesem Skill in
docs/skill-learnings.md→ Sektion „karpathy-skills".
These four principles are inspired by Andrej Karpathy's X-post on LLM coding pitfalls (2026-01-26). They were authored by Jiayuan Zhang (@forrestchang) and contributors under the multica-ai organization, in the upstream repo multica-ai/andrej-karpathy-skills under MIT license.
Karpathy himself does not maintain Claude Code skills. Any external references attributing this skill to Karpathy directly are mis-attributions.
This is a one-page behavioral checklist, not an execution framework. It pairs well with task-driven execution skills (see "Relation to other goal-driven frameworks" below). Reach for it when:
Tradeoff: These guidelines bias toward caution over speed. For trivial tasks, use judgment.
Don't assume. Don't hide confusion. Surface tradeoffs.
Before implementing:
Minimum code that solves the problem. Nothing speculative.
Ask yourself: "Would a senior engineer say this is overcomplicated?" If yes, simplify.
Touch only what you must. Clean up only your own mess.
When editing existing code:
When your changes create orphans:
The test: Every changed line should trace directly to the user's request.
Define success criteria. Loop until verified.
Transform tasks into verifiable goals:
For multi-step tasks, state a brief plan:
1. [Step] → verify: [check]
2. [Step] → verify: [check]
3. [Step] → verify: [check]
Strong success criteria let you loop independently. Weak criteria ("make it work") require constant clarification.
The fourth principle, Goal-Driven Execution, overlaps in spirit with GSD (/gsd-*) and Superpowers' executing-plans workflow. The boundary is scope:
| Layer | This skill | GSD / Superpowers |
|---|---|---|
| What it is | A one-page behavioral checklist (4 principles) | A full execution framework (spec → plan → execute → verify) |
| When to invoke | Inline reminder during a coding task | At the start of a feature or refactor |
| Output | Mental adjustment | Files in .planning/, atomic commits, verification reports |
| Pairs with | GSD, Superpowers, any task framework | This skill (as the inline reminder during execute-phase) |
In other words: GSD and Superpowers (/gsd-plan-phase, /superpowers:executing-plans) structure what you build and how the work moves. These four Karpathy-inspired principles structure how you write each line while you do it. Use them together, not instead of each other.
If you find yourself reaching for these principles in place of a planning workflow on a non-trivial task, prefer the planning workflow — these guidelines are not a substitute for a plan with verifiable steps.
Provides behavioral guidelines to reduce common LLM coding mistakes, focusing on simplicity, surgical changes, assumption surfacing, and verifiable success criteria.
Searches, retrieves, and installs Agent Skills from prompts.chat registry using MCP tools like search_skills and get_skill. Activates for finding skills, browsing catalogs, or extending Claude.
npx claudepluginhub lars-hh/lars-cc-skills --plugin karpathy-skills