Why Your CLAUDE.md Is Making Claude Worse, and the Prompt That Audits It

Anthropic removed over 80 percent of Claude Code's system prompt for the Claude 5 generation and measured no loss on their coding evaluations. Their explanation was blunt: they had been over-constraining Claude Code, through the system prompt and through their own CLAUDE.md files and skills. When they read back real transcripts they found single requests carrying instructions that contradicted each other, one telling Claude to leave documentation as appropriate and another telling it not to add comments.

Your setup has the same problem for the same reason. Every rule you wrote was written for a model that needed it. Most of them were written for models that are two generations old. On Opus 5 they are not guardrails any more, they are noise the model has to read, weigh and resolve before it does anything.

This guide is what actually changed, how to tell which of your rules still earn their place, and a prompt you can paste into Claude that does the audit for you.

• Verified: Anthropic cut over 80 percent of Claude Code's system prompt, announced 24 July 2026
• Their word for the old approach: over-constraining
• The fix is not deleting everything. It is keeping the small number of rules that still bind.

People describe this differently but it is one failure with three faces. Every one of them is the model spending its attention on your instructions instead of your problem.

1. 1
   Answers take longer. Before Claude writes anything it has to read your CLAUDE.md, load the skills that matched, and work out which instruction wins where two of them overlap. That resolution is real work and you pay for it on every single turn.
2. 2
   Simple rules get ignored. This one feels like the model getting worse and it is the opposite. When twenty instructions are in context and three of them conflict, following any one of them means breaking another. A rule that is always in tension with a neighbour is a rule the model learns to discount.
3. 3
   Hallucinations you were not seeing before. Long, prescriptive instruction blocks push the model toward satisfying the instructions rather than checking the world. Anthropic's own framing is that newer models have better judgment and can handle these decisions without explicit guardrails, and that older models needed the constraints to avoid producing the wrong thing.

Six shifts, published by Thariq Shihipar on the Claude blog. They are useful to you because they are a description of what a Claude 5 generation model no longer needs to be told.

1. 1
   Rules become judgment. Explicit constraints like "default to writing no comments" were replaced by higher-level guidance and the model was left to decide.
2. 2
   Examples become interface design. Instead of showing the model how to use a tool, make the tool's interface obvious.
3. 3
   Everything upfront becomes progressive disclosure. Verification and review instructions load when they are needed, as skills, rather than sitting in context on every turn.
4. 4
   Repetition becomes tool descriptions. Guidance that was repeated in three places moved into the description of the tool it was about, once.
5. 5
   Manual memory becomes auto-memory. Saving context by hand with the hash key was replaced by memory the model maintains itself.
6. 6
   Simple specs become rich references. Plain markdown descriptions were replaced by things with more signal in them: real code, test suites, rubrics, working artifacts.

The pattern under all six is the same. Stop describing what you want in prose and start curating what the model sees.