Anthropic runs an ablation on the Claude Code prompt at every model release, and ships the switch that lets you copy it

Read from the on-disk transcript rather than watched. The captions are auto-generated, so everything below is paraphrase cited to a timestamp, and nothing is quoted.

The method is at 5:55 to 6:31. On a new model, delete the whole system prompt, then bring it back one line at a time to find out what each line is worth. Cherny calls this an ablation and describes it as an eval run backwards. At 4:38 to 5:19, he describes the operator-visible version: a flag to set your own system prompt, and an undocumented environment variable he describes as claude code simple set to 1, which strips every system prompt including the tool prompts. Anthropic uses it internally as the ablation harness. The exact spelling is not recoverable from an auto-generated transcript and is not guessed here, so check it against the CLI. His finding from running it, at 5:06, is that the model is slightly more intelligent without the prompts, and that what remains is there to make the product behave rather than to make the model think.

Two structural claims follow. At 6:24 he says almost all the code left in the Claude Code harness is safety, permissions, static analysis and UI. At 20:29 he names verification as the single thing practitioners most often get wrong. Evals get a shelf life too, at 9:37 to 10:25: one to three model generations before saturation forces a rewrite. The advice at 6:37 to 7:15 includes ordinary users as well as harness builders. Every six months, delete your CLAUDE.md, your skills and your hooks, and see what happens.

Four claims in the talk sit in the Dark Factory sweep's unverified queue at confidence 0.20 to 0.55 and are not reported here as fact: prompt injection no longer demonstrable against Opus 5, Bun rewritten Zig to Rust by one steered prompt over 11 days, a Swift rewrite still running after two weeks, and 20 to 30 standing routines doing the work of dozens or hundreds of engineers. Cherny bounds his own headline at 30:14: coding is solved for the kind of coding he does, not for deep systems code, distributed systems or pixel-level UI verification.

Why it matters: The ablation is reproducible tonight and costs one environment variable. If you have been accumulating a CLAUDE.md for a year, this is a first-party argument that much of it is dead weight written for a model that no longer needs it, plus a method for finding out which part. Treat the 80% as an anecdote about Anthropic's prompt; treat the delete-and-restore loop as the transferable thing.