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-meta-reflectiongit 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-meta-reflection)<a href="https://agentmods.dev/skills/learnprompt/andrej-karpathy-skills/karpathy-meta-reflection"><img src="https://agentmods.dev/badge/skills/learnprompt/andrej-karpathy-skills/karpathy-meta-reflection/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-meta-reflection"><img src="https://agentmods.dev/badge/skills/learnprompt/andrej-karpathy-skills/karpathy-meta-reflection.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
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 | $0.01627 |
| Opus 5 | $0.00050 | $0.00813 |
| Sonnet 5 | $0.00020 | $0.00325 |
| Haiku 4.5 | $0.00010 | $0.00163 |
Grade A, and why
karpathy-meta-reflection 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 12d 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 — 184 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Skill 10: Meta-Reflection（元反思）
Source: https://x.com/karpathy/status/2004607146781278521 | https://x.com/karpathy/status/2015883857489522876 "I've never felt this much behind as a programmer" — 56k likes
Core Principle
The landscape is shifting faster than your intuition updates. Deliberate reflection is how you stay calibrated.
Karpathy tracks three things relentlessly:
- What skills are atrophying (AI is taking over, for better or worse)
- What's exploding (new leverage you didn't have before)
- What's your real job now (has the role description changed under your feet?)
Monthly Career Audit Prompt
Run this once a month, 20-30 minutes:
Monthly AI Impact Audit — [MONTH YEAR]
My role/field: [YOUR JOB OR DOMAIN]
Part 1: ATROPHYING SKILLS
Things I've been having AI do in the last month:
[LIST 5-10 TASKS YOU DELEGATED TO AI]
For each, analyze:
- How much would it hurt if AI wasn't available for this?
- Is my ability to do this manually getting worse?
- Is that acceptable? (YES = commodity skill / NO = must maintain)
Part 2: EXPLODING LEVERAGE
New things I can now do (or do much faster) because of AI:
[LIST 3-5 NEW CAPABILITIES OR SPEED MULTIPLIERS]
For each: what specifically unlocked this? What should I double down on?
Part 3: ROLE REDEFINITION
Complete this: "In 2023, my job was to [X]. In [CURRENT YEAR], my job is actually to [Y]."
Part 4: NEXT MONTH FOCUS
1. One skill I will DELIBERATELY practice without AI:
2. One new AI capability I will DELIBERATELY learn:
3. One thing I will STOP doing manually because AI is better:
The Atrophy/Leverage Matrix
For any skill you have, plot it on this 2x2:
HIGH LEVERAGE WITH AI
│
MUST RETAIN │ MAXIMIZE
(ai makes it faster │ (your competitive
but you need the │ edge lives here)
judgment) │
│
HIGH ──────────────────────┼────────────────────── LOW
ATROPHY │ ATROPHY
│
SAFELY DELEGATE │ ABANDON/AUTOMATE
(commodity, ai is │ (low value, losing
better, let it go) │ it doesn't matter)
│
LOW LEVERAGE WITH AI
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.
- 12d ago First seen · 184 lines · 100 tokens per session scan A 2b8172cd96af
karpathy-meta-reflection is a skill published in the GitHub repository LearnPrompt/andrej-karpathy-skills (97 stars, last pushed 2mo ago), licensed MIT. It adds 100 tokens to every session and 1,627 once invoked, about $0.0005 per session on Opus 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.
Other skills, from other repositories
live-tutorial
Create an annotated MP4 tutorial from a workflow demonstrated in the owner's actual application. Use when asked for a tutorial video, not merely to watch an action live.
ArXiv
Search and retrieve arXiv academic papers by topic, category, or paper ID — with AlphaXiv-enriched AI-generated overviews. Uses arXiv Atom API across cs.AI/cs.LG/cs.CL/cs.CR/cs.MA/cs.SE/cs.IR. Three workflows: Latest, Search, Paper. USE WHEN arxiv, papers, latest papers, research papers, recent ML papers, paper…
Teach
Interactive voice-narrated tutorial on whatever the current conversation is about — teaches the topic beat by beat with ASCII diagrams in the terminal while speaking the actual lesson content aloud, asking one question per beat and quizzing until the ideas land. USE WHEN teach me, /teach-me, tutorial, teach me this…
ghostex-help
Use this skill whenever the user asks how Ghostex works, what a Ghostex feature, view, button, or setting does, how to do something in Ghostex ("how do I", "where is", "can Ghostex"), or asks you to change the app for them ("make the terminal match the chat width", "move the sidebar", "turn off the completion sound"…
read-book
When you want to read and extract structured notes from a book — PDF, EPUB, MOBI, markdown, .txt, pasted text, or URL to a public-domain work. Reads in chunks (by chapter when a TOC exists, by 50-page blocks otherwise), extracts per-chapter TL;DR + key concepts + quotes + action items + frameworks, and offers to…
subscription-setup
A setup workflow for creating recurring medical-learning subscriptions after a doctor’s profile has been registered.