Breaking

Anthropic Urges Claude Code Users to Audit Old Prompts for Claude Opus 5.5

October 3, 2026 at 07:15 EDT

Software Dev & Coding
AI Agents
Foundation Models

Anthropic is turning prompt cleanup into a recommended migration step for Claude Opus 5.5: Claude Code users can run /claude-api prompt-audit to find older instructions that may now add cost, latency or rigid behavior. The guidance targets CLAUDE.md, skills, agents, commands, tool descriptions and application prompts that were written for less capable models.

September 2026 | Anthropic

Old Claude prompts now need an audit

Claude Code's prompt-audit commands check long-lived project instructions for patterns that can add cost, latency or rigid behavior on Claude Opus 5.5.

36%

higher cost from legacy prompts in a support-desk test

14%

lower cost after running the audit on the same prompts

token context window for Claude Opus 5.5

$4/$20

per million input and output tokens

Measured lift in the support-desk evaluation

Accuracy rose after audit while cost moved down.

92%Before audit

97%After audit

The cleanup loop

Open Claude Code

→

Run prompt-audit

→

Review findings

→

Apply chosen diff

Why developers like it

It points to exact files and lines, then proposes a diff instead of asking teams to rebuild prompt surfaces by hand.

What still needs judgment

The audit is not an automatic deletion pass; useful constraints should stay, and clean prompts should produce no patch.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.