What's this about?
Thariq, a Claude Code core developer, revealed that the team removed roughly 80% of Claude Code's system prompt for the newest Claude 5 models, and wrote up what they learned about writing system prompts, skills, and CLAUDE.md files. The core lesson: stronger models need less instruction, and piles of defensive rules that older models required now actively make newer models worse.
What exactly is a "system prompt"?
A system prompt is the standing instruction sheet an AI product silently feeds its model before your first message — who it is, what tools it has, what it must never do. Think of it as an employee handbook taped inside the AI's locker. For years the industry coped with model weaknesses by making the handbook longer; this is a major product saying the handbook had become the problem.
What's the community saying?
- Validation of a trend: less scaffolding, more model — echoing Uncle Bob's "I don't read agent code" from Friday as part of the same shift of trust toward the model.
- Practical takeaway hunting: people mining the article for what to delete from their own CLAUDE.md files.
- Skeptics noting the irony that the same weekend, Redditors found a hardcoded "don't use subagents" instruction for Opus 5 still in the prompt — pruning is incomplete.