Skip to content
AI Dev Toolkit.
Esc
  • Loading…

Claude Code · Requests and models

Claude Code is unable to respond to this request (Usage Policy)

Why Claude Code says a request “appears to violate our Usage Policy” or “can’t help with this”, why it repeats, and how to get the session working again.

Message

The exact error

Claude Code
API Error: Opus 4.6 can't help with this. Start a new session to continue.

Send feedback with /feedback or learn more: https://www.anthropic.com/legal/aup

Also appears as:

Before v2.1.219
Claude Code is unable to respond to this request, which appears to violate our Usage Policy (https://www.anthropic.com/legal/aup). Please double press esc to edit your last message or start a new session for Claude Code to assist with a different task.
With no model recorded
API Error: Claude can't help with this. Start a new session to continue.
Reasoning-extraction refusals add
Details: `[reasoning_extraction]`

Meaning

What does “appears to violate our Usage Policy” mean?

The API declined to answer because something in the conversation triggered a check against Anthropic’s Usage Policy. The check reads the whole conversation, not just your last prompt, so the same session usually keeps refusing even after you change the subject, and so does a session reopened with --continue or --resume, because the transcript still holds the content. Current versions name the model that declined (“Opus 4.6 can’t help with this”) and include a Request ID and Message ID you can quote to support. It isn’t a ban or an account action by itself.

Causes

Why did my request trigger it?

  • Content earlier in the conversation, such as a pasted file, a web page or a tool result, that matches a policy check
  • A prompt worded in a way the check reads as harmful, even when the task is legitimate
  • An instruction asking Claude to write out its reasoning verbatim, in a CLAUDE.md, skill, subagent or output style (the message then adds Details: [reasoning_extraction])
  • On Amazon Bedrock, Google Cloud and Microsoft Foundry, a request flagged as a cybersecurity topic, which shows this message too

Fix

How do I get past a Usage Policy refusal?

  1. Rewind to before the triggering turn

    Press Esc twice or run /rewind, pick the message before the refusal, then rephrase or take a different approach. This removes the triggering content from what’s sent next.

  2. If you can’t tell which turn caused it, clear

    Run /clear to start a fresh conversation in the same project. The old one stays on disk and in /resume, but resuming it brings the refusal back.

  3. In -p scripts, start a new session

    Rewind isn’t available in non-interactive mode. Retry with a rephrased prompt in a new session, without --continue. Policy checks vary by model, so trying another model with --model can help in some cases.

  4. For “[reasoning_extraction]”, remove reasoning-dump instructions

    Delete any instruction to write a <thinking> section, a scratchpad or a reasoning field, from your prompt or from CLAUDE.md, skills, subagent prompts, output styles and MCP tool descriptions. Run claude --safe-mode to test with customisations off, then start a new session after editing. Asking for a short explanation of the answer is fine.

  5. Report a wrong refusal

    Run /feedback and quote the Request ID and Message ID from the message.

Don’t do this

  • Rewording the same request again and again in the same session: the earlier content is still sent, so it keeps failing.
  • Resuming the old session with --continue or --resume: the transcript still contains the triggering content.

Explained

How is it different from the other refusal messages?

Claude Code has three content-related errors. This one is a policy check on the conversation. “<model>’s safeguards flagged this message” is a separate cybersecurity check on the Anthropic API, with a link to Anthropic’s Cyber Verification Program for legitimate security work. Output blocked by content filtering policy is a filter on Claude’s output when it starts reproducing existing text, such as a licence file. Each has its own fix, so read the exact wording first.

If the refusal came from instructions in your project files, the CLAUDE.md generator and the CLAUDE.md guide help you write short, plain instructions that are less likely to set it off.

FAQ

Questions people ask

Why does Claude Code keep refusing after I change the subject?

The Usage Policy check reads the whole conversation, not only your newest message, so the content that triggered it is still being sent. Rewind to the turn before the refusal with Esc twice or /rewind, or run /clear to start a fresh conversation. Resuming the old session brings the content back.

Will a Usage Policy refusal get my account banned?

A single refusal is the API declining one request, not an account action. Anthropic’s Usage Policy does allow it to act on accounts that violate the policy, so don’t keep pushing the same request. If you believe the refusal is wrong, report it with /feedback and include the Request ID shown in the message.

What does “Details: [reasoning_extraction]” mean?

Safeguards read the request as asking Claude to reproduce its internal reasoning, for example a required thinking or scratchpad section. Remove that instruction from your prompt and any CLAUDE.md, skill, subagent prompt or output style, test with claude --safe-mode, then start a new session.

Does switching models fix a Usage Policy refusal?

Sometimes. Anthropic’s error reference says policy checks vary by model, so another model may answer. In an interactive session, rewind past the triggering turn first; in -p mode, start a new session with a rephrased prompt and, if needed, a different --model.

Related

Search all Claude Code errors

Keep exploring