A few random notes from Claude coding quite a bit last few weeks
TL;DR Highlight
Andrej Karpathy shares honest observations from weeks of coding with Claude — productivity gains, brain atrophy, 'sleepwalking', and workflow tips.
Who Should Read
Developers actively using or considering AI coding tools, especially those worried about skill degradation or over-reliance.
Core Mechanics
- Karpathy reports real productivity gains — tasks that would take hours complete in minutes. The speed multiplier is real and not just hype.
- He coins 'sleepwalking' to describe a failure mode: going along with AI-generated code without truly understanding it, shipping things you can't debug or maintain.
- Brain atrophy is a genuine concern he names explicitly: if you stop writing code yourself, you lose fluency. The skill degrades through disuse just like a muscle.
- His recommended workflow: use AI for boilerplate, scaffolding, and unfamiliar APIs — not for core algorithmic logic where deep understanding matters most.
- He still writes critical, novel, or security-sensitive code manually — the AI assists, not replaces, on high-stakes decisions.
- The most useful prompt pattern he found: describe the goal and constraints, let Claude propose an approach, then critically evaluate before accepting — not rubber-stamping.
Evidence
- Karpathy's credibility as an author here is very high — he helped invent the systems he's critiquing, making this unusually honest and well-informed self-reflection.
- HN reaction was strong agreement: many developers independently noticed the 'sleepwalking' phenomenon without having a name for it.
- Counter-arguments were made that brain atrophy concern is overstated — we don't worry about calculators atrophying mental arithmetic, so why worry about AI atrophying code writing?
- The Karpathy response: there's a difference between arithmetic (algorithmic, rule-based) and programming judgment (creative, contextual) — the latter doesn't have a clean calculator analogy.
- Several senior engineers noted they deliberately avoid AI for certain practice tasks to maintain fluency, accepting slower delivery in exchange for skill preservation.
How to Apply
- Audit your current AI coding usage: are there categories of code where you've stopped writing yourself? Deliberately reintroduce manual coding for those categories periodically.
- Before accepting AI-generated code, be able to explain every line: what it does, why it's correct, and what edge cases it handles. If you can't, don't ship it.
- Reserve AI assistance for: library integration, boilerplate, unfamiliar language syntax, test generation. Handle core logic, security decisions, and novel algorithms yourself.
- Set personal projects where AI assistance is deliberately off-limits — maintaining the ability to work without AI is a hedge against dependency.
Terminology
Sleepwalking (AI coding)Karpathy's term for accepting and shipping AI-generated code without truly understanding it — you're moving forward without being fully present in the process.
Brain atrophyThe gradual loss of coding fluency from disuse — if you never write code manually, you lose the ability to do so confidently.
Vibe codingCoding primarily by describing intent to an AI and accepting its output — productive for many tasks but risky for complex or critical code.