Anthropic removed more than 80% of Claude Code’s system prompt and the coding evaluations did not drop.

Not dropped a little. No measurable loss.

Thariq Shihipar of Anthropic published it on July 24. The detail that makes the post worth reading in full is what the problem turned out to be: that prompt had been carefully hand-tuned over months for earlier models, and the care itself was what held Opus 5 and Fable 5 back.

Your rules file aged and nobody noticed

If your team has been running agents for a year, you have an instruction file that grew by sedimentation.

Somebody added a rule because the model did something strange in March. Somebody else pasted three examples of how to call a tool because in April it kept calling it wrong. A third person repeated one instruction in both the prompt and the tool description, just to be safe.

Every one of those decisions was correct on the day it was made.

None of them got revisited when the model changed underneath.

What you end up with is a recipe card taped to the wall telling a twenty-year cook to wait thirty seconds before adding oil. They will read it. They will follow it. And they will cook worse than they know how.

The six shifts, translated

Anthropic lists six. Here they are in a form you can act on.

Rules to judgment: state the goal and the constraints that genuinely matter, then let the model choose the path.

Examples to interface design: if you need three examples for a tool to be used correctly, the tool’s parameters are the problem, not the example count.

Upfront information to progressive disclosure: stop front-loading everything, let context load when it is needed.

Repetition to concise description: the same instruction in the prompt and the tool definition does not reinforce, it confuses.

Manual memory to auto memory: less hand-maintenance of memory files, more automatic capture of what matters.

Simple specs to rich references: code, test suites and rubrics teach more than paragraphs of Markdown.

An independent re-measurement put the actual cut nearer 70% than 80%, because some memory blocks were made conditional rather than deleted outright. The conclusion holds either way.

This contradicts what most of us were doing

The uncomfortable half. Through 2025 the prevailing practice, ours included, was to write more. More context, more examples, more explicit rules, more scaffolding around the model.

It worked, because the models of that period needed the scaffolding.

Now the scaffolding is the ceiling. And because instruction files never fail loudly, nobody audits them. The agent keeps running, keeps shipping, and you never find out it has been working below its capability for six months.

We covered where context actually breaks in a running operation in context is the enterprise AI bottleneck, with the fuller framework in context engineering for AI agents.

How I would audit it

Run the doctor command on your project. It exists for this and takes a minute.

Then do the manual pass, which is the one that teaches you something. Take your instruction file and ask of each rule: did we add this for a problem that still exists? You will find rules written for a model you no longer use, a tool you already replaced, a failure mode that stopped happening in February.

Delete them. Not comment them out. Delete.

Then measure. Same twenty tasks before and after. If quality holds, you just recovered context budget and speed for free. If it drops, add back exactly one rule, the one that was load-bearing, and measure again.

That loop takes an afternoon and almost nobody runs it, because removing things feels risky and adding things feels productive. Anthropic just published the evidence that it is the other way around.

Let’s audit the rules you gave your agents