Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add LearnPrompt/andrej-karpathy-skills --skill karpathy-agentic-engineeringgit clone --depth 1 https://github.com/LearnPrompt/andrej-karpathy-skillsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/learnprompt/andrej-karpathy-skills/karpathy-agentic-engineering)<a href="https://agentmods.dev/skills/learnprompt/andrej-karpathy-skills/karpathy-agentic-engineering"><img src="https://agentmods.dev/badge/skills/learnprompt/andrej-karpathy-skills/karpathy-agentic-engineering/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/learnprompt/andrej-karpathy-skills/karpathy-agentic-engineering"><img src="https://agentmods.dev/badge/skills/learnprompt/andrej-karpathy-skills/karpathy-agentic-engineering.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector warn
SkillSpector: 3 findings, up to medium
These are SkillSpector’s own severities. On a checked sample its high-severity flags on skills were ~96% false positives — a documented command, a public API, a “never do X” rule — so we show them as a caution to read, not a verdict. Why →
- medium Memory Poisoning · line 29 Skill attempts to fill the context window with filler content, displacing legitimate instructions and safety constraints. This can degrade agent performance or bypass safety boundaries.Fix: Implement context-window management that detects and rejects padding or stuffing attempts. Prioritize system instructions over user-injected content.
- medium Excessive Agency · line 81 Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.Fix: Add human-in-the-loop confirmation for destructive, irreversible, or high-impact operations. Never auto-execute commands that modify files, send data, or alter system state.
- medium Excessive Agency · line 94 Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.Fix: Add human-in-the-loop confirmation for destructive, irreversible, or high-impact operations. Never auto-execute commands that modify files, send data, or alter system state.
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00100 | $0.01279 |
| Opus 5.5 | $0.00040 | $0.00512 |
| Sonnet 5.5 | $0.00020 | $0.00256 |
| Haiku 4.5 | $0.00010 | $0.00128 |
Grade A, and why
karpathy-agentic-engineering scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2mo ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 129 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Skill 1: Agentic Engineering（代理工程）
Source: https://x.com/karpathy/status/2004607146781278521 | https://x.com/karpathy/status/2015883857489522876 | https://x.com/karpathy/status/2026731645169185220 "It is hard to communicate how much programming has changed." — ~37k likes
Core Principle
You are no longer writing code. You are directing a team of tireless junior engineers.
The shift: write code → write tasks with success criteria → agent loops → human taste review
The 7-Step Inner Loop (Karpathy's Incremental Coding Rhythm)
This is the micro-loop for every coding session:
- Stuff context — Load entire project into context (
files-to-prompt, relevant docs, recent diffs) - Discuss plan — Ask agent for high-level approaches with pros/cons before any code
- Request first draft — Pick one approach, get a minimal first implementation
- Inline learning — Manually review the diff; ask agent to explain any API you don't recognize
- Test — Run tests, check outputs, verify success criteria
- Commit — Only commit what you understand and have verified
- Ask for next — Let agent suggest the next incremental step
Never skip step 4. Understanding atrophies if you rubber-stamp every diff.
CLAUDE.md / AGENTS.md Template
Write this file at the project root to prevent the most common agent failure modes:
# Agent Instructions for [Project Name]
## Behavior Rules
- Make one incremental change at a time. Do NOT refactor unrelated code.
- Always explain WHY you're making a change before the diff.
- If unsure about intent, ASK before implementing.
- Do NOT add new dependencies without explicit approval.
- Prefer simple, readable solutions over clever ones.
## Success Criteria (task-specific — fill in per session)
- [ ] Tests pass: `[test command]`
- [ ] No new lint errors: `[lint command]`
- [ ] Behavior matches: [description of expected output]
## Project Context
- Stack: [language, framework, key libs]
- Entry point: [file]
- Test command: [command]
- Forbidden: [list things agent must never do]
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2mo ago First seen · 129 lines · 100 tokens per session scan A 5382bbbc7f4d
karpathy-agentic-engineering is a skill published in the GitHub repository LearnPrompt/andrej-karpathy-skills (122 stars, last pushed 3mo ago), licensed MIT. It adds 100 tokens to every session and 1,279 once invoked, about $0.0004 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
karpathy-agentic-engineering compared with skills from other repositories
| Mod | Type | Stars | Last push | Licence | Scan | Tokens | Works with |
|---|---|---|---|---|---|---|---|
| karpathy-agentic-engineeringLearnPrompt/andrej-karpathy-skills | Skill | 122 | 3mo ago | MIT | A | 100 | Claude Code |
| imagegenopenai/codex | Skill | 134k | today | Apache-2.0 | A | 113 | Codex |
| playwright-climicrosoft/playwright | Skill | 99k | today | Apache-2.0 | A | 19 | Claude Code |
| local-ai-agentsmicrosoft/ai-agents-for-beginners | Skill | 79k | today | MIT | A | 200 | Claude Code, Codex |
| systematic-debuggingobra/superpowers | Skill | 314k | 3d ago | MIT | A | 21 | Claude Code |
| next-cache-components-adoptionvercel/next.js | Skill | 143k | today | MIT | A | 95 | Claude Code |
| next-partial-prefetching-adoptionvercel/next.js | Skill | 143k | today | MIT | A | 103 | Claude Code |