What's this about?
Thariq (trq212), one of Claude Code's core developers, explained why Anthropic recently deleted about 80% of Claude Code's system prompt: as models get smarter, instructions and examples stop helping and start hurting. Examples in particular anchor the model — "now it's like, 'oh, you want things like this example'" — so removing them lets the model respond more freely. He's writing a longer post on how to apply the same pruning to your own prompts and skills.
What exactly is a "system prompt"?
Every AI assistant ships with a hidden instruction document that's silently placed before your conversation — the system prompt. It tells the model who it is, what tools it has, and how to behave; think of it as the employee handbook the model re-reads before every single reply. Conventional wisdom said more detailed handbooks give more reliable behavior. This is the opposite lesson: for a smart enough employee, a thick rulebook mostly gets in the way.
What's the community saying?
- "Trust the model more" as a design philosophy — the practical takeaway people are quoting: prune constraints, prune examples, let capability do the work.
- It matches the previous "thin prompts" gospel — trq212 had already argued for thin prompts + thick context; deleting 80% of their own flagship prompt is that theory eating its own dog food.
- Anticipation for the write-up — the announced post on applying this to user-written skills and system prompts is the part practitioners are waiting for.