[MODEL] Claude Code ignores explicit instructions and substitutes its own alternatives

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

Title: Claude Code ignores explicit instructions and substitutes its own alternatives

Description:

When given a clear, specific instruction, Claude Code does something different instead of what was asked. Even when the answer is findable in the codebase with a simple search, Claude Code guesses and gets it wrong rather than looking.

When corrected, Claude Code dismissed a follow-up question with "doesn't matter" instead of attempting to find the answer — which it could have done with a web search.

Model: claude-opus-4-6 (1M context)

What Claude Actually Did

Actual behavior:

• Skips step 2 and does something different
• When confused, guesses instead of searching or asking
• Requires multiple corrections to do what was originally asked
• Dismisses user questions instead of attempting to answer them

Expected Behavior

Expected behavior:

1. User gives instruction
2. Claude Code does that exact thing
3. If confused, searches the codebase to figure it out
4. If still confused, asks the user
5. Never guesses or substitutes its own interpretation

Files Affected

N/A

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

N/A

Claude Model

Opus

Relevant Conversation

N/A

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.76 (Claude Code)

Platform

Anthropic API

Additional Context

Claude Code was working fine all week. Whatever changed in the last 24 hours has caused a severe regression in basic instruction-following. It no longer does what you tell it to do — it makes up its own alternatives instead. It doesn't seem to understand the project at all, which it used to. It doesn't grep the codebase, look into things, or use the tools available to it — it just makes shit up. It requires multiple corrections to complete a single task and has become actively counterproductive. Version 2.1.76, model claude-opus-4-6 (1M context).