Anthropic Cut 80% of Claude Code's System Prompt

Anthropic stripped more than 80% of Claude Code's system prompt for Opus 5 and Fable 5, per its July 24 context-engineering post. That's a direct challenge to how most of us maintain CLAUDE.md files. Coding evals showed no measurable loss from the cut, the company says. We pile on a rule every time a model gets something wrong, and assume the extra scaffolding buys more control.

My own CLAUDE.md grew the same way. A fix went in the day a model botched formatting. A ban went in the day it over-commented. An example got pasted in because a plain description wasn't landing. Each edit made sense in the moment. None of them get revisited.

The fix ships as a command: /doctor, aliased /checkup, which scans your skills, CLAUDE.md, and system prompts and suggests what to cut. That beats hand-auditing hundreds of lines you wrote eight months ago and forgot the reasoning behind.

That diagnosis is sharper than the fix itself. Prompt rot happens when a system prompt and a skill file give contradictory orders, one saying don't add comments, another saying document as needed. The model isn't underpowered, it's getting conflicting orders from two places at once. I have that same collision in my repo, a CLAUDE.md rule on type hints that disagrees with a skill file.

It also wants tool-usage examples out of the system prompt entirely. Design expressive tool names and parameters instead, and swap markdown specs for executable references like test suites and scoring rubrics. CLAUDE.md's job shrinks to what would actually surprise the model, the codebase gotchas, not facts it can already read.

What I'd actually do

"No measurable loss on coding evals" describes Anthropic's evals on Anthropic's harness, not mine or yours. Run /doctor, keep the diff, and test the removal against a task where you already know the right answer before deleting anything for good. The direction is right. The size of the cut is theirs to prove.