The team behind Claude Code deletes most of Claude Code's prompt every time a new model ships. I think it's the most useful thing they've said this year.
Boris Cherny, who leads Claude Code, said on stage that for Opus 5 they cut more than 80% of the system prompt. With each new model they run an ablation: delete the whole prompt, bring it back line by line, keep only what changes behaviour.
The harness was never the product. It's something you re-derive for every model.
Last week I tried to automate that re-derivation at YAGNI, where agent teams take engineering tickets to a merged pull request. The plan: let a self-improving harness loop (Harness²) propose edits to our agent's skill and prompts, test them on real tickets, keep what wins.
I froze it after the first smoke runs. Every failure we found (a model's chatty preamble becoming a commit message, 18 minutes spent looking for a browser the sandbox didn't have, a verification step overrunning its clock) was an environment or code bug, found by reading logs. None of them was a prompt edit a search would discover. And one full run costs $2–5 and ~30 minutes, so separating a real improvement from noise takes dozens of runs per variant.
What we do instead: save every run's transcript, cut it at the turn where things went wrong, and replay from there with the current skill and with the proposed edit. Cents and minutes instead of dollars and half an hour, and each edit is tested at the exact decision it's meant to change.
My takeaway: before you let a model rewrite your harness, read your logs. Most of what looks like a prompt problem isn't one.
And if your CLAUDE.md has been growing since last year, it's probably encoding fixes for a model you no longer use.