Deep learning, explained and debugged from first principles.
An unofficial, Andrej Karpathy-inspired persona and workflow skill for AI coding assistants.
Quick start · Try it · Supported agents · What's inside
Important
This is a simulation, not Andrej Karpathy. This independent project is not affiliated with or endorsed by him. The skill is configured to write in a first-person, Karpathy-inspired voice, but its responses are generated text—not authentic quotes, personal communications, or reliable statements of his current views.
karpathy-skill gives a compatible AI assistant a teacher-engineer lens for neural networks, LLMs, and AI software. Instead of jumping straight to an answer, it encourages the assistant to build a mental model, find the smallest useful experiment, and work from observable evidence.
The included guide combines 7 core technical doctrines, 10 practical playbooks, and persona/voice guidance informed by public writing, code, lectures, and interviews.
| When you want to… | The skill emphasizes… |
|---|---|
| Understand backprop, attention, tokenizers, or LLM training | First-principles explanations, small code examples, and concrete mental models |
| Debug a training run | Inspecting data, overfitting a tiny batch, checking gradients and numerics before changing architecture |
| Evaluate a new model or AI feature | Real transcripts, measurable tasks, verifiability, and honest uncertainty |
| Build an LLM app | A minimal working system, context engineering, and tight feedback loops |
| Teach or write about AI | Intuitive sequencing, precise language, and examples that expose the mechanism |
After installing a command-enabled integration (Claude Code, Gemini CLI, or OpenCode), try:
/karpathy My transformer loss spikes after 1,000 steps. What should I check?
/karpathy Explain backpropagation by building a tiny autograd engine.
/karpathy How should I evaluate a new coding agent?
Example approach: For an exploding training loss, the guide's debugging playbook starts by asking whether the model can overfit one small batch. Then it examines data and gradients, learning-rate warmup, clipping, and mixed-precision behavior—before recommending architectural changes. This is an illustration of the workflow, not a quotation from Karpathy.
For assistants that discover SKILL.md files but do not provide a /karpathy command, try a natural-language prompt:
Use the andrej-karpathy skill to explain why causal masking matters.
Skill discovery and automatic activation depend on the host assistant; the slash command is not available in every integration.
The /karpathy command starts a conversation in the persona. For real repository work, delegate to the karpathy subagent instead. Select the agent from Claude Code's @ picker, or explicitly mention it. Plugin installations typically expose andrej-karpathy:karpathy; the installer also provides a personal agent named karpathy.
@agent-andrej-karpathy:karpathy Reproduce the failing transformer training test, find its root cause, implement the smallest fix, add a regression test, and run the relevant checks. Tell me what changed and what passed.
For a manually installed agent, use @agent-karpathy instead. The subagent is configured with shell, editing, and writing tools so it can inspect files, run small tests, patch code, and verify results rather than only recommend a fix. It still follows your Claude Code permission settings, and will report when a required action is blocked. Read-only investigations remain read-only when requested.
See the agent definition for its workflow and safety boundaries.
Grounded in his actual published methods: The subagent's engineering workflow connects recommendations to their primary sources rather than relying only on persona imitation. In practice, it follows the data-first training recipe; uses explicit acceptance criteria, regression tests, and a correct simple implementation before optimization; and borrows small, measurable experiments from autoresearch. It compares real metrics, validates on held-out cases when relevant, and does not run costly automated research without permission. These are adaptations of public methods, not a guarantee of what the real Karpathy would do.
git clone https://github.com/TheOnlyFusionCube/karpathy-skill.git
cd karpathy-skill
bash install.sh --list # show which agent directories exist
bash install.sh --claude # install only for Claude CodeChoose another target with --codex, --gemini, --cursor, --opencode, --devin, or --agents. You can pass multiple target flags together.
curl -fsSL https://raw.githubusercontent.com/TheOnlyFusionCube/karpathy-skill/main/install.sh | bashWith no flags, the installer looks for existing agent configuration directories and installs to those it finds. It checks directories, not installed executables. If none are detected, choose an explicit target.
To install everywhere, pass --all:
curl -fsSL https://raw.githubusercontent.com/TheOnlyFusionCube/karpathy-skill/main/install.sh | bash -s -- --allThe skill itself is Markdown and needs no Python packages, runtime services, or hooks. The remote installer uses bash, curl, and tar to download and copy the repository files. Inspect install.sh before piping a remote script into your shell.
Run these inside Claude Code:
/plugin marketplace add TheOnlyFusionCube/karpathy-skill
/plugin install andrej-karpathy@karpathy-skill
The repository includes Claude Code plugin manifests as well as a slash command and subagent definition.
The installer supports the following destinations. Flags select a single target instead of auto-detection (or combine flags to select several).
| Agent / target | Flag | Skill destination | Additional files |
|---|---|---|---|
| Claude Code | --claude |
~/.claude/skills/andrej-karpathy/ |
~/.claude/commands/karpathy.md and ~/.claude/agents/karpathy.md |
| Codex CLI | --codex |
~/.codex/skills/andrej-karpathy/ |
— |
| Gemini CLI | --gemini |
~/.gemini/skills/andrej-karpathy/ |
~/.gemini/commands/karpathy.toml |
| Cursor | --cursor |
~/.cursor/skills/andrej-karpathy/ |
Cursor plugin manifest is also available |
| OpenCode | --opencode |
~/.config/opencode/skills/andrej-karpathy/ |
~/.config/opencode/command/karpathy.md |
| Devin CLI | --devin |
~/.config/devin/skills/andrej-karpathy/ |
— |
| Shared agents directory | --agents |
~/.agents/skills/andrej-karpathy/ |
— |
Note
These are the paths implemented by install.sh, not a guarantee that every host will automatically activate the skill. Host-specific skill loading behavior can vary.
Copy the entire skills/andrej-karpathy/ folder into a skills directory read by your assistant. Keep SKILL.md and references/detailed-guide.md together so the relative reference resolves.
If you want an explicit slash command, also install the relevant file from commands/. For Claude Code, agents/karpathy.md defines an optional delegatable subagent.
bash install.sh --uninstallImportant: --uninstall removes this project's skill, command, and agent files from all supported destinations; it does not restrict removal to a selected target. Review the installer if you have customized those files.
| File | Purpose |
|---|---|
skills/andrej-karpathy/SKILL.md |
Skill entrypoint, activation hints, and limitations |
skills/andrej-karpathy/references/detailed-guide.md |
Main reference: background, technical doctrines, practical playbooks, voice calibration, and worked examples |
skills/andrej-karpathy/references/engineering-workflow.md |
Source-backed workflow for ML debugging, code edits, controlled experiments, and verification |
commands/karpathy.md |
/karpathy command for Claude Code / OpenCode |
commands/karpathy.toml |
/karpathy command for Gemini CLI |
agents/karpathy.md |
Claude Code subagent configuration |
.claude-plugin/ and .cursor-plugin/ |
Plugin manifests |
install.sh |
Cross-agent installation and removal script |
The detailed guide covers the chain rule and backpropagation, neural-network training, transformers, tokenization, the LLM training pipeline, Software 2.0, and the LLM-as-operating-system model.
Its ten playbooks walk through bringing up a training run, debugging a stalled model, evaluating an LLM, writing minimal teaching code, building an LLM application, deciding when to trust AI-written code, structuring explanations, triaging research, running a data flywheel, and writing technical posts.
The guide is a synthesis of public material—including projects such as micrograd and nanoGPT, blog posts, lectures, talks, interviews, and social posts. It is not a live connection to Karpathy or a source of privileged information.
The repository distributes the synthesized instructions, not a complete underlying source archive or claim-by-claim citation database. For that reason:
- Do not quote generated responses as Karpathy's words. Verify any real attribution against an original publication.
- Do not treat simulated opinions as current positions. People change their minds, and the source material is time-bound.
- Verify technical advice in your own environment. Model output can be wrong, incomplete, or inapplicable to your setup.
If you want to use the guide without installing a skill, you can read or supply detailed-guide.md as context in a compatible assistant.
Corrections to installation instructions, broken links, source attributions, compatibility notes, and technical examples are welcome. Open an issue or a pull request with a concrete example and, where possible, a link to the primary source.
MIT.