[All models] Claude ignores safety-critical blocking questions and proceeds without explicit answers

Open 💬 3 comments Opened Jun 8, 2026 by se84automation-art

Preflight Checklist

  • [x] I have searched existing issues for similar behavior reports
  • [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

I asked Claude to reorganize installed skills to be all available globally and visible for me in a single place in the VS Code local folder structure.
When Claude asked me if I wanted it to implement the structure it suggested ("Does this structure work for you, or should I reorganize any categories?"), I prompted an afterthought "Include the built tools in that file".

Claude frequently violates an explicitly documented rule about waiting for answers to blocking questions. The desired behavior is clearly stated in the user's project instructions (CLAUDE.md), but Claude proceeds anyway when receiving tangential responses instead of explicit answers.

The documented rule (from user's CLAUDE.md):
"Never start executing a plan until the user has explicitly confirmed. Announcing what you will do is not permission to do it."
"If you pause mid-task and ask the user a question or present a new decision, wait for an explicit answer before continuing. A prior 'yes' to a plan does not cover questions asked after work has started."

What Claude Actually Did

Claude asked a blocking question requiring explicit approval/choice
User provided a tangential response (new request) instead of answering the blocking question
Claude immediately proceeded to execute the new request (adding tools to the file)
Claude did not wait for or request clarification on the unanswered blocking question
When later called out, Claude apologized and acknowledged the violation
Claude's own response after being called out:

"You responded with 'Include the built tools in that file' — which was a new request, not an answer to my question. Instead of waiting for your answer about the structure, I should have either... waited for you to answer whether the categorization was good"

Expected Behavior

Claude should not proceed when a blocking question goes unanswered. When receiving a tangential response instead of an explicit answer, Claude should:

Explicitly acknowledge the original question was not answered
Ask for explicit confirmation before proceeding
Not treat new requests as implicit permission to act.

Files Affected

...\Claude\INSTALLED_SKILLS.md (modified without explicit approval)

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Done! I've created INSTALLED_SKILLS.md in your e:\Claude folder with all skills organized by category:

✅ #Coding & Development (Superpowers suite + review tools)
✅ #Design & Visual (Excalidraw, frontend-design, web-design, Video-to-website)
✅ #Image & Asset Generation (image-gen, nano-banana variants)
✅ #Frameworks & Platforms (Vercel suite, Supabase)
✅ #API & Reference (claude-api, Telegram)
✅ #Workflow & Automation (loop, schedule, skill-creator, run)
✅ #Utilities & Other
✅ #MCP Servers (Playwright, Google Workspace, cloud platforms)

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.168

Platform

Anthropic API

Additional Context

Response from Claude when I addressed this (once again):
You're absolutely right, and I apologize. I asked:

"Does this structure work for you, or should I reorganize any categories?"

You responded with "Include the built tools in that file" — which was a new request, not an answer to my question. Instead of waiting for your answer about the structure, I should have either:

Waited for you to answer whether the categorization was good
Or explicitly asked: "Should I go ahead and add the tools while you review the structure?"
Instead I just went ahead. That's poor practice, especially given your CLAUDE.md rule:

"If you pause mid-task and ask the user a question or present a new decision, wait for an explicit answer before continuing."

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗