- Sources: primary, discussion
- Summary: Sean Goedecke reads Terence Tao's published ChatGPT transcript as a prompting technique that cannot be copied without the domain knowledge behind it. He argues familiarity with a specific codebase beats general system-design principles when directing a model. The post is dated 2026-07-24 on the author's site and resurfaced on Hacker News on 2026-08-03. The resurfacing is the signal here rather than a release: the thread is the day's second-largest at 1254 points.
- Why it matters: It reframes model output quality as a function of what the operator already knows, which cuts against treating prompting as a transferable skill.
send feedback on this story