David Crawshaw, a former staff engineer at Google and co-founder of Tailscale, published a detailed personal account of how he uses large language models while programming. The post, which hit the front page of Hacker News with 919 points and 332 comments, catalogs three modes: autocomplete, search, and what he calls ‘chat-driven programming.’

Crawshaw estimates that for every two hours of programming he accepts more than ten autocomplete suggestions, uses an LLM for a search-like task once, and engages in a chat session once. He describes the shift as comparable to the 1995 moment he first had continuous internet access — ‘astonishing, and felt like the future.’

The post dives into the craft of writing prompts, advocating for ‘exam-style questions’ — providing specific objectives and all background material while avoiding messy IDE contexts. Crawshaw also argues that LLMs alter code engineering tradeoffs: because they excel with small, isolated contexts, developers should prefer smaller packages and more tests, even if that means writing more code overall. He concludes that fixing the mistakes in LLM-generated code is often easier than starting from scratch.

The Hacker News thread is split. Some commenters — several with senior credentials — report trying the tools and finding them useless for their daily tasks, despite open-mindedness. Others say Crawshaw’s credibility makes the endorsement especially meaningful, comparing the current moment to earlier tooling rifts over CNC machines and IDEs. A smaller number of commenters invoke literary warnings, comparing the trend to E.M. Forster’s ‘The Machine Stops.’