They Deleted 80% of Their AI Instructions. The Model Did Better.

The team behind Claude Code deleted 80% of the instructions they give their AI. The remaining 20% turned out to be the only part that was actually useful. Claude Code is a programming tool from Anthropic. Its creator, Boris Cherny, said recently that every time a new model comes out, the team follows the same process: delete all the AI instructions, then add them back one by one to see which ones actually help. When they tested this with their latest model, Opus 5, they ended up keeping just 20%. Cherny went even further and removed everything entirely, running the model with no instructions at all. Without any guidance, it performed slightly better. The instructions that got cut were written a year ago to work around what the model could not do back then: handle this kind of problem this way, try that approach when you run into the other issue. The model needed those reminders. The newer model reads a task and knows what to do. The old instructions had become friction, pulling the model away from its own judgment and back onto outdated paths. This is uncomfortable because it goes against instinct. When AI does not do what you want, the natural response is to add more explanation: break down the steps, give more context, name every exception. A software engineer with twenty years of experience said he used to review every line of code that AI wrote for him. A single set of changes can run thousands of lines, so thorough review was not really possible anyway. Now he watches two things: whether the external interface or the data structure changed in unexpected ways. If those are fine, he lets the rest go. He said the hardest part was accepting that less oversight actually worked better. More explanation is usually not the fix. You are applying a mental model built around AI from six months ago. The model has moved ahead. Cherny's advice is specific: every few months, delete your AI instructions and settings and see how the model handles things without them. The list you wrote for the previous version might not be needed at all anymore. Next time you give AI a task, try starting with the goal and how you will know when it is done. Leave the rest out. See what comes back.