15 Engineers Share How They Actually Use AI to Code

Gregor Ojstersek compiled workflows from 15 senior engineers and CTOs on how they use Claude Code, Cursor, and Copilot in production. The strongest consensus: spend up to 50% of project time on design and specs before letting AI generate anything, and never trust the first output. One contributor merges SEV2 fixes in 27 minutes start to finish. Another ships solo what would normally require a small team. The recurring anti-pattern is rubber-stamping AI-generated PRs as velocity increases - multiple contributors insist on adversarial testing with a separate agent from the one that wrote the code. The emphasis on CLAUDE.md files and spec-driven development echoes what Anthropic's own roundtable with Stripe and NVIDIA engineers surfaced weeks ago - the teams getting the most from AI coding tools are the ones investing heavily in context before prompting.