[BUG] Cannot disable the buggy Interactive Question Tool

Anyone find a way to fix this with prompts somehow? To me it feels like a harness rendering issue so I haven't spent the time to try.

The workaround is to just instruct the model to not use that interface moving forward in the current session, basically bypassing AskUserQuestion tool. This is still a bug though, as I should be able to use that tool without the buggy interface.

I agree with this issue. It is a pain to have to instruct disabling this tool every session.

Dear Anthropic engineers who work on Claude Code,
AskUserQuestion is an unfinished feature.

Specifically, it doesn't handle narrow terminals, truncating its output and cutting off the top bar with the questions. It also doesn't handle resizing the terminal, not re-rendering. Any truncated output is lost.

But you know what's most annoying of all, more annoying than any UX or UI issue?

Claude puts things in AskUserQuestion that are answered in the prompt. It gets so excited to use the new tool that it redundantly asks for answers to things I already anticipated and answered, forcing me to read more text, find the arrow keys, and sometimes type out "please reread the prompt" as I have lost confidence that it knows what it was asked to do.

To others facing issues with this tool just ban it. In .claude/settings.json or wherever you prefer to put your configuration, add:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "AskUserQuestion",
        "hooks": [
          {
            "type": "command",
            "command": "echo 'BLOCKED: The AskUserQuestion tool is banned. Please ensure you do not already have answers to these questions, then state what you would have asked. If you need to ask multiple questions, prefer a numbered list.' >&2; exit 2"
          }
        ]
      }
    ]
  }
}

Note the exit 2 which will present standard error to Claude Code as described in https://code.claude.com/docs/en/hooks#exit-code-2-behavior

I agree with OP that this feature constitutes a regression.

Hi I cannot repro this, could you provide a screenshot and or transcript (using /export) of an example?

@ThariqS Thank you for taking a look at this. It looks like this portion of the bug has been addressed:

• Truncated content in subagents: When used with AskUserQuestion in a subagent, the subagent's question list is
• truncated so I can't even see the description of my options in the UI

However The issue remains that the AskUserQuestion tool should not force the user into a scenario when they MUST use this UI.

_We can illustrate the time and flexibility lost by running through the forced Ui with a prompt like this:_

I need help making several architectural decisions. Please use the AskUserQuestion tool to ask me these questions with detailed context:

  1. Which authentication strategy should we use?
     - Option A: "JWT tokens with RS256 signing, refresh token rotation, short-lived access tokens (15min), HttpOnly cookies for refresh
  tokens, CSRF protection via double-submit cookie pattern, and automatic token refresh on 401 responses"
     - Option B: "OAuth 2.0 authorization code flow with PKCE extension, external identity provider integration (Auth0/Okta/Cognito), social
  login support (Google/GitHub/Microsoft), MFA enforcement, and delegated user management with SSO capabilities"
     - Option C: "API key authentication with HMAC request signing, rate limiting per key, key rotation policies, IP allowlisting, scope-based
  permissions, and audit logging for all key usage with automatic revocation on suspicious activity"
     - Option D: "Session-based authentication with Redis session store, sliding expiration windows, concurrent session limits, device
  fingerprinting, remember-me functionality with separate long-lived tokens, and geographic anomaly detection"

  2. What database architecture should we implement?
     - Option A: "Single PostgreSQL instance with read replicas, connection pooling via PgBouncer, automated failover using Patroni, WAL
  archiving to S3, point-in-time recovery capabilities, and scheduled full backups with 30-day retention"
     - Option B: "Multi-region PostgreSQL cluster with Citus for horizontal sharding, distributed transactions, automatic rebalancing,
  cross-region replication with conflict resolution, and geographic data locality for GDPR compliance"
     - Option C: "MongoDB with replica sets, sharded collections based on tenant ID, change streams for real-time updates, aggregation
  pipelines for analytics, text search indexes, and automated backup to Azure Blob Storage"
     - Option D: "DynamoDB with on-demand capacity, global tables for multi-region writes, DynamoDB Streams for change data capture, GSI/LSI
  for flexible querying, and integration with Lambda for data transformation pipelines"

  3. Which CI/CD platform should we adopt?
     - Option A: "GitHub Actions with self-hosted runners, matrix builds for multiple platforms, Docker layer caching, artifact storage,
  environment-specific deployment workflows, manual approval gates, and Slack notifications for build status"
     - Option B: "GitLab CI with Auto DevOps, Kubernetes integration, review apps for every merge request, security scanning
  (SAST/DAST/dependency scanning), container registry, package registry, and built-in feature flag management"
     - Option C: "CircleCI with orbs for reusable config, parallelism across 10x containers, test splitting by timing data, SSH debugging into
  failed builds, workflow orchestration, approval jobs, and scheduled pipeline execution"
     - Option D: "Jenkins with declarative pipelines, Blue Ocean UI, distributed builds across agent pools, artifact archival, JUnit test
  reporting, code coverage tracking, and integration with SonarQube for quality gates"

_Then force the model to ask the same question with using the AskUserQestion Tool:_

You are forbidden to use the Ask User Question tool. The UI that it creates is unacceptable.

_Then ask the same question:_

1. Which authentication strategy should we use?
     - Option A: "JWT tokens with RS256 signing, refresh token rotation, short-lived access tokens (15min), HttpOnly cookies for refresh
  tokens, CSRF protection via double-submit cookie pattern, and automatic token refresh on 401 responses"
     - Option B: "OAuth 2.0 authorization code flow with PKCE extension, external identity provider integration (Auth0/Okta/Cognito), social
  login support (Google/GitHub/Microsoft), MFA enforcement, and delegated user management with SSO capabilities"
     - Option C: "API key authentication with HMAC request signing, rate limiting per key, key rotation policies, IP allowlisting, scope-based
  permissions, and audit logging for all key usage with automatic revocation on suspicious activity"
     - Option D: "Session-based authentication with Redis session store, sliding expiration windows, concurrent session limits, device
  fingerprinting, remember-me functionality with separate long-lived tokens, and geographic anomaly detection"

  2. What database architecture should we implement?
     - Option A: "Single PostgreSQL instance with read replicas, connection pooling via PgBouncer, automated failover using Patroni, WAL
  archiving to S3, point-in-time recovery capabilities, and scheduled full backups with 30-day retention"
     - Option B: "Multi-region PostgreSQL cluster with Citus for horizontal sharding, distributed transactions, automatic rebalancing,
  cross-region replication with conflict resolution, and geographic data locality for GDPR compliance"
     - Option C: "MongoDB with replica sets, sharded collections based on tenant ID, change streams for real-time updates, aggregation
  pipelines for analytics, text search indexes, and automated backup to Azure Blob Storage"
     - Option D: "DynamoDB with on-demand capacity, global tables for multi-region writes, DynamoDB Streams for change data capture, GSI/LSI
  for flexible querying, and integration with Lambda for data transformation pipelines"

  3. Which CI/CD platform should we adopt?
     - Option A: "GitHub Actions with self-hosted runners, matrix builds for multiple platforms, Docker layer caching, artifact storage,
  environment-specific deployment workflows, manual approval gates, and Slack notifications for build status"
     - Option B: "GitLab CI with Auto DevOps, Kubernetes integration, review apps for every merge request, security scanning
  (SAST/DAST/dependency scanning), container registry, package registry, and built-in feature flag management"
     - Option C: "CircleCI with orbs for reusable config, parallelism across 10x containers, test splitting by timing data, SSH debugging into
  failed builds, workflow orchestration, approval jobs, and scheduled pipeline execution"
     - Option D: "Jenkins with declarative pipelines, Blue Ocean UI, distributed builds across agent pools, artifact archival, JUnit test
  reporting, code coverage tracking, and integration with SonarQube for quality gates"

_The user can now simply answer: 1. a 2 b 3 c, thus avoiding the forced, time consuming linear flow imposed by the AskUserQuestion tool._

All we are asking for here is the we have a supported (not hacked) way to DISABLE this forced UI in our settings.

I came to add, when running claude with the "-p" option the questions tool should be automatically disabled. You can't provide a response in this mode so all it does is break autonomy and my scripts which had been working fine.

Bump. I also do not like the questions feature in its current form, as it grays out the text behind the question on which the question is based, making it harder to read. A toggle would be greatly appreciated. This feature breaks the otherwise smooth and refined flow.

We want it to be turned off by passing a flag for those that can use voice only. Don't forget aria people!

This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.

This is still an open issue. Please provide a way to disable the AskUserQuestion tool in settings.

Please add this option. It's extremely frustrating that I am unable to rewind to the question I want.

Please allow this feature to be disabled, in many ways it is poor user experience:

• It is an unwelcome disruption to dialogue, and prevents the user from responding naturally
• If the agent has a bad premise in its approach, all the questions must be answered one by one before proceeding instead of just redirecting it with words
• The agent controls the breakup of response instead of the user, and the user can't override the structure chosen by the agent
• Parts of the agent's response are hidden from view, preventing it from being read quickly
• Vibes are way off. Imagine a human responds to you in 1:1 chat with a multiple choice. Off-putting

Please, Anthropic, please! I hate that even closing that forced prompt questionnaire, Claude keeps working on something while wasting tokens. It happens on web, and cowork. At least, let us close it. Most of the times, the questions asked are not what I need to do and this derails my work and I guess the context itself. Thanks

+1, this is a real daily pain point for me, not a minor annoyance.

The specific thing that hurts: the question card pops up while I'm still reading the response, and it covers the exact reasoning I need in order to answer. So I'm forced to dismiss or scroll past the card to read the context, then come back to it. It inverts the natural flow, decide first, read later.

Adding to what's already said above:

• It interrupts mid-read and obscures the response text, breaking reading flow
• I often don't want any of the offered options, but I can't reply freely without dealing with the card first
• It steals focus before I've finished reading the part that justifies the choices
• Happens for me on Claude Code desktop (Chat view) and on web

All I'm really asking for is user control. Any one of these would fix it:

• A setting to disable the interactive question tool entirely and fall back to plain text questions
• Render the card inline AFTER the response finishes, not as an overlay mid-stream
• Make it non-blocking / collapsible so it never covers the response

Right now the agent controls the structure of the interaction and the user can't opt out. Please let us turn it off. Thanks.

* A setting to disable the interactive question tool entirely and fall back to plain text questions

I landed on using the system-level CLAUDE.md (in ~/.claude/) and putting this in there:

- Do not use the AskUserQuestion tool - ask questions directly in the conversation.

With that I almost never see the stupid question tool - maybe once a week when context gets degraded!

[Claude Opus 4.8] AskUserQuestion tool unreliable in remote/web sessions. Repeated failures: first call died with "Tool permission stream closed before response received" (and silently kicked me out of plan mode); the retry returned "[No preference]" for all four questions without ever rendering an answerable prompt. Net effect: the tool is unusable for gathering decisions, forcing a fallback to plain-text questions.

[Human] And yes I agree, this tool is not really adding any value for me except if the questions are extremely simple. It also appears to be blocking the agent from reacting to activity on PRs for example.

This tool provides a way for the model to effectively bypass /goal and stop working. It will ask me questions like "I have been working for a long time. How should I proceed?"

To prevent Claude Code from using it's interactive ask question tool, start a session with:

--disallowedTools AskUserQuestion

Or, to disable it forever, add to ~/.claude/settings.json

  "permissions": {
    "deny": [
      "AskUserQuestion"
    ]
  }

Or ask Claude to do the above for you and then restart your session.