How do I install the Prompts Karpathy skill?

Run npx skillmds@latest add andersonlimahw/prompts-karpathy in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Prompts Karpathy safe to use?

This skill has not completed SkillMD's automated safety review yet. Capability flags: docs only. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Prompts Karpathy?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Prompts Karpathy free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Prompts Karpathy?

Andersonlimahw (@andersonlimahw) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Prompts Karpathy by andersonlimahw — Agent Skill

Name: Prompts Karpathy
Rating: 4.666666666666667 (3 reviews)
Author: Andersonlimahw

File contents

Prompts — Karpathy method

The trio karpathy-guidelines (behavioral principles), karpathy-recipe (recipe for medium/large work), and karpathy-loop (autonomous optimization cycles). Complementary, not redundant: guidelines apply always; recipe structures large work; loop chases a metric.

R1 — Guidelines on any implementation

Skills: karpathy-guidelines When: any non-trivial code — the default to cut classic LLM mistakes.

Implement <feature/fix> in <path> applying karpathy-guidelines:
state assumptions before coding (ask if one is structural), the
minimum code that solves it, surgical changes (no "improving"
neighboring code), and a numbered plan with "→ verify:" per step.
Success: every line of the diff traces to the request; every
verify ran with evidence.

R2 — Full recipe for medium/large work

Skills: karpathy-recipe + verification-before-completion When: a feature/refactor that spans several files or sessions.

Structure <work> with karpathy-recipe: break it into stages with a
verifiable criterion each, execute stage by stage, and close each
one with verification-before-completion before moving on.
Success: plan with stages + verification; no stage marked complete
without the check having run; final diff with no extra scope.

R3 — Review against the four principles

Skills: karpathy-guidelines + code-review-expert When: reviewing a diff/PR looking for the vices the guidelines prevent.

Review the diff of <branch/PR> with code-review-expert using
karpathy-guidelines as the rubric: (1) hidden assumptions,
(2) speculative complexity/premature abstraction, (3) non-surgical
changes outside scope, (4) weak or unverified success criteria.
Success: every finding mapped to a principle, with file:line and a
1-sentence fix.

R4 — Autonomous metric loop

Skills: karpathy-loop When: an objective metric to optimize in cycles. Full recipe in track prompts-feedback-loops (R4) — one loop driver at a time.

---name: prompts-karpathydescription: Ready-to-paste prompts for the Karpathy method — behavioral guidelines while coding, the full recipe for medium/large work, review against the four principles, and the autonomous optimization loop.---# Prompts — Karpathy methodThe trio `karpathy-guidelines` (behavioral principles), `karpathy-recipe` (recipe for medium/large work), and `karpathy-loop` (autonomous optimization cycles). Complementary, not redundant: guidelines apply always; recipe structures large work; loop chases a metric.## R1 — Guidelines on any implementation**Skills:** `karpathy-guidelines`**When:** any non-trivial code — the default to cut classic LLM mistakes.```textImplement <feature/fix> in <path> applying karpathy-guidelines:state assumptions before coding (ask if one is structural), theminimum code that solves it, surgical changes (no "improving"neighboring code), and a numbered plan with "→ verify:" per step.Success: every line of the diff traces to the request; everyverify ran with evidence.```## R2 — Full recipe for medium/large work**Skills:** `karpathy-recipe` + `verification-before-completion`**When:** a feature/refactor that spans several files or sessions.```textStructure <work> with karpathy-recipe: break it into stages with averifiable criterion each, execute stage by stage, and close eachone with verification-before-completion before moving on.Success: plan with stages + verification; no stage marked completewithout the check having run; final diff with no extra scope.```## R3 — Review against the four principles**Skills:** `karpathy-guidelines` + `code-review-expert`**When:** reviewing a diff/PR looking for the vices the guidelines prevent.```textReview the diff of <branch/PR> with code-review-expert usingkarpathy-guidelines as the rubric: (1) hidden assumptions,(2) speculative complexity/premature abstraction, (3) non-surgicalchanges outside scope, (4) weak or unverified success criteria.Success: every finding mapped to a principle, with file:line and a1-sentence fix.```## R4 — Autonomous metric loop**Skills:** `karpathy-loop`**When:** an objective metric to optimize in cycles. Full recipe in track [prompts-feedback-loops](../prompts-feedback-loops/SKILL.md) (R4) — one loop driver at a time.## See also- `verification-before-completion` — universal "is this actually done?" gate.- `continual-learning` — persist lessons from the cycles across sessions.

Frequently asked questions

Prompts Karpathy