The old prompt engineering playbook is falling apart. Thariq Shihipar, a member of the Claude Code team, says his group stripped over 80% of the system prompts from Opus 5 and Fable 5 — and coding benchmarks barely budged.
In the past, the team packed Claude Code with prohibitions and examples to stop the model from randomly deleting files or writing excessive comments. But newer models can already judge the situation, blending user requests with code style. Pile on too many rules, and they start contradicting each other.
Shihipar's advice: keep your CLAUDE.md lean. Just include a project overview and the genuinely tricky spots. Break out complex processes like testing and code review into Skil... (the post cuts off, but presumably into separate skill files).