karpathy-ideas
llms.txt1 starsChanged yesterdayMITView on GitHub
Unofficial, curated notes on Andrej Karpathy's public ideas about working with LLMs and AI agents. Each idea: date, source link, summary in our own words, one short quote at most, and practical Apply steps (views, his opinions and explanations, may have none). Apply steps are our translation of his idea into practice, not his words; a step marked "Our reading" or "our addition" goes further than the idea itself, and one "Our reading" line at the top of Apply marks…
In this repository4 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- llms-full.txtllms.txt
- karpathy-ideasSkill
The file
We split this llms.txt into the rules it gives an agent
2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 5
Not affiliated with or endorsed by Andrej Karpathy.
- Distinctive: line 26
Use the themes when you need ideas for one task (a prompt, a workflow, a tool), not a project review.
≈ “Common” is an estimate: at least 1 in 5 of the 3,727 llms.txt 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 Ideas > Unofficial, curated notes on Andrej Karpathy's public ideas about working with LLMs and AI agents. Each idea: date, source link, summary in our own words, one short quote at most, and practical Apply steps (views, his opinions and explanations, may have none). Apply steps are our translation of his idea into practice, not his words; a step marked "Our reading" or "our addition" goes further than the idea itself, and one "Our reading" line at the top of Apply marks all its steps. Not affiliated with or endorsed by Andrej Karpathy. Prefer ideas with `status: current`; superseded ideas link to his newer view. Quotes belong to their author: link to the source instead of reproducing more text. *(Anthropic period)*: dated 2026-05-19 or later, after he [joined Anthropic](https://x.com/karpathy/status/2056753169888334312), and about Claude or Anthropic. Weigh these with that in mind. ## Applying the ideas to a project To review a whole project, use the [project checklist](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/checklist.md), not the themes: 17 questions about agent setup and token use, how the project is built, and the product. Each item gives his idea with the link, what it looks like in a project, questions to ask with what to measure, and the trade-offs his ideas name; proposals are options for the project's owner. Measure each question first (where it says "Ask"), then open only the linked ideas that point to a gap. Read the files from disk, not through a web-fetch tool. Download them with `curl` (for example `curl -sLO https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/checklist.md`) or `git clone https://github.com/ntroccoli/karpathy-ideas` into a temporary folder outside the project: this file, checklist.md, corrections.md, log.md and any idea file you open. A web-fetch tool may hand you a smaller model's summary instead of the text (Anthropic calls Claude Code's WebFetch "lossy by design": https://code.claude.com/docs/en/tools-reference#webfetch-tool-behavior). Delete the folder when the work with it is done. First look for the project's record: `docs/karpathy-ideas.md`, or the record file the last review named. It decides the path: - **No record: a first review.** Measure every checklist question; do not judge from memory. Write numbered proposals in the report format (step 4 of the checklist's "How to run this checklist"), and change nothing. - **A record: a revisit.** Read [corrections.md](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/corrections.md) first (errors in what this collection told projects to do, and how to repair them), then the [log.md](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/log.md) entries newer than the one the record names. Re-check the recorded items, re-measure what the checklist's "On a revisit" names, and propose only what is new, earlier "later" items and items that stopped holding. A skipped or declined proposal is closed: raise it again only for a new idea or a changed measurement, and say which. Either way, say in one line which path you took and why, and change nothing until the user chooses. The record is written when they do (the checklist's step 8). Whoever ran the review between 2026-11-09 and 2026-11-15 should run the review prompt again: because a record exists, it does the revisit ([correction C1](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/corrections.md#c1-size-limits-of-ours-2026-11-09-to-2026-11-15)). No warranty: the ideas, the checklist and its steps are provided as they are. Show the user every proposal and wait for their choice before changing anything, suggest a branch or a backup first, apply one proposal at a time and stop at the first failing check, and remember the maintainer is not liable for what applying them changes. ## For a single task: how to read this cheaply Use the themes when you need ideas for one task (a prompt, a workflow, a tool), not a project review. 1. Pick the 2-3 themes below that match your task and read only those theme pages (one line per idea). Each page lists practices (ways of working) first, then views (his opinions and explanations); for a task, the practices are usually enough. Read the full index only if you need everything. 2. Open only the idea files you will use. Each is short: YAML front matter (title, date, url, kind, summary, themes, status and the other fields), then a `# Title` heading, a summary paragraph with no heading of its own, `## In his words` (optional, one short quote), `## Apply` (steps; views may have none) and `## Context`. 3. If you can delegate, let a smaller, cheaper model or subagent do steps 1-2 and return a shortlist of idea links with one line each on why they fit. The main model then reads only the shortlisted files. This saves cost and keeps the main context clean. Relative links inside these files resolve against https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/ (for example `../ideas/x.md` in a theme page is https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/ideas/x.md). ## Themes - [agentic-coding](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/themes/agentic-coding.md): Coding with agents: workflows, IDEs, review, vibe coding vs agentic engineering. - [verification](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/themes/verification.md): Success criteria, tests, verifiability, eval loops, autoresearch. - [context-and-prompting](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/themes/context-and-prompting.md): Context engineering, prompting, stating scope, giving intent. - [output-and-reading](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/themes/output-and-reading.md): Output formats, simplified English, HTML, diagrams, reading with LLMs. - [knowledge-management](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/themes/knowledge-management.md): Personal wikis, notes, compiling knowledge with LLMs. - [agent-native-software](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/themes/agent-native-software.md): Software 3.0, LLM OS, CLIs and docs for agents, products. - [llm-psychology](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/themes/llm-psychology.md): Jagged intelligence, ghosts vs animals, sycophancy, how to model LLMs. - [security-and-dependencies](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/themes/security-and-dependencies.md): Supply chain, sandboxing, dependency hygiene. - [learning-and-education](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/themes/learning-and-education.md): Learning with LLMs, teaching, skill atrophy. - [research-and-training](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/themes/research-and-training.md): Training, RL, tokenizers, small models, research practice. - [ai-progress](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/themes/ai-progress.md): Views on the field, timelines, society, jobs. ## Reference - [Index](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/index.md): every idea, one line each, newest first (larger than a theme page) - [Agent instructions](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/AGENTS.md): file format, themes, how to read and maintain ## Optional - [All ideas in one file](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/llms-full.txt): for tools that cannot browse. About 340 KB (~87k tokens), so prefer the themes when you can follow links - [Sweep log](https://raw.githubusercontent.com/ntroccoli/karpathy-ideas/main/log.md): what was collected and when
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.

