karpathy-methodology
Skill91 starsChanged 6 months agoMITView on GitHub
Apply Andrej Karpathy AI methodology and principles from his 2023-2026 insights. Use this skill when the user wants to apply Karpathy-style thinking, needs guidance on agentic engineering, LLM knowledge bases, minimalist coding, understanding-first principles, or any of the 14 core methodologies distilled from his work. Also triggers on: "karpathy method", "karpathy approach", "karpathy way", "AK style", "like karpathy does", "karpathy principles", or when user asks how Karpathy would approach any programming or AI problem.
In this repository14 other files it gives its agents.
- karpathy-agentic-engineeringSkill
- karpathy-autoresearchSkill
- karpathy-education-firstSkill
- karpathy-idea-filesSkill
- karpathy-llm-simulatorSkill
- karpathy-llm-wikiSkill
- karpathy-meta-reflectionSkill
- karpathy-minimalismSkill
- karpathy-output-evolutionSkill
- karpathy-practice-environmentsSkill
- karpathy-supply-chain-hygieneSkill
- karpathy-system-prompt-learningSkill
- karpathy-understanding-firstSkill
- karpathy-vibe-to-agenticSkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 51
Use AI to go faster. Use your brain to know if you're going in the right direction.
- Distinctive: line 81
When the user hasn't specified a sub-methodology, use this all-in-one Karpathy-style framing:
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Karpathy Methodology — 14 Core Skills
Distilled from Andrej Karpathy's most-liked posts on X (2023–2026), his LLM Wiki Gist, and the multica-ai/andrej-karpathy-skills repo. Source: https://x.com/karpathy | https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
This skill is the master index. When invoked, either:
- Apply the specific sub-methodology relevant to the user's current task, or
- Guide the user to the right one from the 14 below.
The 14 Methodologies
| # | Skill Name | One-line Essence |
| 1 | karpathy-agentic-engineering | Orchestrate agents with clear tasks and verifiable success criteria |
| 2 | karpathy-llm-wiki | Let LLM maintain your knowledge base; you do the exploring |
| 3 | karpathy-llm-simulator | Ask LLM to simulate expert debate, not give one opinion |
| 4 | karpathy-minimalism | 200 lines of pure Python beats 50 npm packages |
| 5 | karpathy-vibe-to-agentic | Vibe coding raises the floor; agentic engineering raises the ceiling |
| 6 | karpathy-autoresearch | Agent loops on git branches; you only change the prompt |
| 7 | karpathy-output-evolution | Text → Markdown → HTML → rendered; always demand structure |
| 8 | karpathy-understanding-first | Outsource thinking, never understanding |
| 9 | karpathy-idea-files | Ship the Gist (abstract + criteria), not the code |
| 10 | karpathy-meta-reflection | Monthly audit: what's atrophying, what's exploding |
| 11 | karpathy-supply-chain-hygiene | Every dependency is an attack surface |
| 12 | karpathy-education-first | Make everything teachable; nano-projects over monoliths |
| 13 | karpathy-system-prompt-learning | Write strategy into the system prompt like a textbook |
| 14 | karpathy-practice-environments | Build gyms where agents can try, fail, and learn |
Meta-Principle (the one that unifies all 14)
"You can outsource your thinking but you cannot outsource your understanding." — Karpathy, ~46k likes, 2026
Use AI to go faster. Use your brain to know if you're going in the right direction.
How to Apply This Skill
Quick Decision Tree
"I need to build/code something" → Start with #1 Agentic Engineering (clear task + success criteria), layer in #4 Minimalism (avoid bloat), finish with #8 Understanding First (verify what was built).
"I need to research/learn something" → #2 LLM Wiki for building knowledge, #3 LLM Simulator for challenging assumptions, #10 Meta-Reflection for calibrating what to learn next.
"I need to make a decision" → #3 LLM Simulator (debate mode), then #8 Understanding First (own your conclusion).
"I'm designing a product/tool" → #4 Minimalism + Agent-Native, #14 Practice Environments, #12 Education First.
"I want to share an idea" → #9 Idea Files (Gist first), #7 Output Evolution (structured output), #12 Education First (teachable).
"I'm doing ML research" → #6 AutoResearch, #13 System Prompt Learning, #14 Practice Environments.
Master Prompt Template
When the user hasn't specified a sub-methodology, use this all-in-one Karpathy-style framing:
You are operating under the Karpathy Methodology.
Core constraints:
1. UNDERSTAND before outsourcing — never ship what you can't explain
2. MINIMIZE dependencies — prefer 200-line pure implementations
3. AGENT-NATIVE outputs — CLI-friendly, markdown-structured, LLM-legible
4. VERIFIABLE goals — every task must have a testable success criterion
5. TEACH as you build — outputs should be understandable by a curious beginner
Task: [USER_TASK]
Success criteria: [WHAT_DONE_LOOKS_LIKE]
Constraints: [TECH_STACK, SIZE_LIMIT, NO_LIBS]Source Posts (Top by Likes)
- 145k likes — Joins Anthropic: https://x.com/karpathy/status/2056753169888334312
- 59k likes — LLM Knowledge Bases: https://x.com/karpathy/status/2039805659525644595
- 56k likes — "Never felt this behind as a programmer": https://x.com/karpathy/status/2004607146781278521
- 46k likes — "Outsource thinking not understanding": https://x.com/karpathy/status/2049907410303865030
- 40k likes — Claude coding notes: https://x.com/karpathy/status/2015883857489522876
- 37k likes — Programming phase shift: https://x.com/karpathy/status/2026731645169185220
- 31k likes — LLM argue the opposite: https://x.com/karpathy/status/2037921699824607591
- 28k likes — AutoResearch project: https://x.com/karpathy/status/2030371219518931079
- 28k likes — litellm supply chain attack: https://x.com/karpathy/status/2036487306585268612
- 25k likes — 243 lines pure Python GPT: https://x.com/karpathy/status/2021694437152157847
- 22k likes — Agent network discussion: https://x.com/karpathy/status/2017442712388309406
- 19k likes — HTML output + I/O evolution: https://x.com/karpathy/status/2053872850101285137
- 27k likes — LLM Wiki Gist version: https://x.com/karpathy/status/2040470801506541998
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.