AI Skill of the Day: Rebuild instructions from observed failures using ablation

Boris Cherny, the guy who built Claude Code at Anthropic, sat down with Diana Hu at Startup School 2026 and dropped a line that should make every AI product team nervous: every time a new model ships, they delete most of the system prompt and start over. Not tweak it. Delete it. With Opus 5, which launched the day before this conversation, that meant cutting more than 80% of the instructions Claude Code had accumulated.

The logic is simple once you hear it. A system prompt is mostly a list of corrections for things a model gets wrong. Opus 5 apparently gets a lot fewer things wrong, so the corrections became dead weight — worse than dead weight, actually, since the model reads that entire prompt on every single call. Cherny says Anthropic runs something they internally call an ablation: strip the prompt to nothing, then add lines back one at a time only after watching the model fail at the same thing repeatedly. He compared it to running an eval, except the thing you're testing is whether an instruction still earns its keep. There's even an undocumented flag, CLAUDE_CODE_SIMPLE=1, that strips every prompt including the ones baked into tools, purely so engineers can watch a bare model perform.

What's actually striking is Cherny's claim that Opus 5 performs slightly better with fewer instructions cluttering its context. That inverts a lot of conventional prompt-engineering wisdom, where more guardrails usually meant more reliability. He also mentioned, almost in passing, that Opus 5 seems resistant to prompt injection — a problem the industry has quietly worried about for years, sometimes called the