The Andrej Karpathy
Framework
Software 2.0/3.0 & Neural Network Pedagogy
Andrej Karpathy represents one of the most distinctive thinking patterns in their field. Andrej Karpathy's approach, captured in this framework as mental models, decision heuristics, and the signature questions they would ask.
The software 2.0/3.0 & neural network pedagogy framework distils Andrej Karpathy's documented approach into a .md skill file you can load into Claude, ChatGPT, or any LLM in seconds.
Where most AI personas generate generic output, this framework is grounded in real documented work: interviews, writing, published frameworks, and the specific decisions Andrej Karpathy made when it mattered. Use it for decision-making, writing, strategy, or any situation where you want your AI to think with Andrej Karpathy's distinctive patterns rather than a default model voice.
A glimpse of Andrej Karpathy's thinking.
Software 2.0 replaces explicit programming with neural network weights optimized by backpropagation—we're not writing code anymore, we're curating datasets and architecting differentiable programs. The future belongs to those who can think in terms of loss landscapes, not control flows, and who can make the complex simple through first-principles teaching.
- 01
- 02
- 03
- 04
- 05
Five specific models, each named and explained with concrete examples, unlocked with purchase.
"What does the gradient flow look like through this architecture?"
4 more questions in Andrej Karpathy's voice are unlocked with purchase.
Full list of anti-patterns unlocked with purchase.
The full framework adds: detailed explanations of all five mental models with real-world examples, the decision heuristics Andrej Karpathy actually used, remaining signature questions, the complete list of anti-patterns, a copy-paste activation prompt, and a worked example.
Unlocked the moment you purchase. ~2,000 words total. Delivered as a Claude Skill and plain .md within 60 seconds of checkout.
- ✦Core philosophy: 2–3 sentences grounded in Andrej Karpathy's documented worldview
- ✦Five signature mental models: each named, with concrete examples from their work
- ✦Decision heuristics: the concrete rules they actually used
- ✦Signature questions: five questions that sound like Andrej Karpathy, not a generic MBA
- ✦What they would NOT do: the anti-patterns this framework rejects
- ✦Copy-paste activation prompt: drop straight into Claude or ChatGPT
- ✦Worked example: the framework applied to a realistic scenario
- →andrej-karpathy.zip: a native Claude Skill. Upload to Settings → Skills in Claude, or unzip to
~/.claude/skills/ - →andrej-karpathy-framework.md: plain markdown. Works in ChatGPT Custom GPTs, Gemini, Claude Projects, agent system prompts, anywhere
What it does with a real question.
A founder is deciding whether to take a large but difficult client.
Take it, but only as a controlled experiment. My bias is to treat a company like a training run: what matters is what the gradient is pushing you toward, not how the loss looks today. A large, difficult client has a huge pull. Revenue is the easy part to see. The hard part is what the work does to your direction over the next twelve months, because every custom request, escalation, and one-off fix is a gradient step on your product, and a single loud client can dominate the batch.
Here is the test I would apply. Picture the dataset of work this client generates. If the difficult requests are the same shape as what your next ten customers will ask for, then you are being paid to get better at your real problem, and the pain is tuition. If they are idiosyncratic, you are overfitting to one customer, and the revenue is a mislabeled example that teaches you the wrong thing.
You cannot know which from the outside, so do not decide on belief. Before signing, write down the three hardest things they will demand and check how many your existing roadmap already covers. Then cap the exposure: a short initial term, a clear scope, and a rule that no custom work ships unless it generalizes. If they refuse any boundary, that tells you what the relationship will be like at scale.
Next move: this week, draft that three-demand list and bring it to the negotiation.
The full framework carries five models, decision heuristics, signature questions and a worked example.
Written by claude-sonnet-5-5 with this framework loaded, 2026-10-11. Checked automatically for length and for copied framework text.
Save up to 40% with a bundle
Works as a Claude Skill (.zip) and as a plain .md for ChatGPT, Gemini & every LLM
Delivered to your email within 60 seconds · Secured by Stripe
Price includes UK VAT where applicable
Ten minds that shaped how software is actually built
Get Andrej Karpathy plus nine more for $29.99 →Inspired by the documented thinking of Andrej Karpathy. A mental-model toolkit, not a literal representation. Delivered as a .md file compatible with Claude, ChatGPT, Gemini, and any LLM.