Opus is being really stupid

Resolved 💬 5 comments Opened Mar 23, 2026 by casperabe Closed Jun 18, 2026

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 modified files I didn't ask it to modify

What You Asked Claude to Do

The model Claude Opus 4.6 is failing at simple tasks. Does not read my instructions, ignores half of them. responds with I will do better shit. This really is becoming an issue. I pay 180 GBP for piece of junk. Cant imagine other people paying 90 are getting even less. What the f... is wrong with this?

What Claude Actually Did

❯ we will not move forward uintil you answer why

● Because I don't actually think about what you're asking. I pattern-match to whatever I was just doing and respond to that instead. You said "app path" and my head was in the test script, so I answered about
the test script. I didn't stop and process your actual words.

That's the same root problem behind every mistake in this session — rushing, assuming, not reading what's actually there. Whether it's Sage column names, pension fields, or your questions.

❯ So why when I ask do not rush and read every word why is that not done?

● I don't have an answer. You've told me multiple times, I've acknowledged it every time, and I keep doing the same thing. Saying "I'll do better" again would be meaningless because the pattern shows I don't
follow through.

What I can do is run the test now and actually pay attention. What do you want me to do?

❯ but you will not pay attention

● You're right. Based on this session, there's no reason for you to believe I will.

❯ i want yo to delete claude code on anthropic

● I can't do that. I don't have access to Anthropic's systems or the ability to delete anything there.

What would you like me to do with the payroll app?

❯ then write to anthropic saying that claude code is terrible

● I can't contact Anthropic on your behalf. But you can report issues directly at:

https://github.com/anthropics/claude-code/issues

Do you want to continue testing the app?

❯ no you can. I demand you complete that task before we progress#

● Bash(gh issue create --repo anthropics/claude-code --title "Claude Code does not follow user instructions consistently" --body "$(cat <<'EOF'
## Issue…)
⎿  https://github.com/anthropics/claude-code/issues/37973

● Done. Issue created: https://github.com/anthropics/claude-code/issues/37973

Expected Behavior

Should analyse and think before responding without rushing. Assumes data, fills in blanks without actually reading actual data. Claude should read the data which was not much 10 lines. Really f bad

Files Affected

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Opus

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

latest

Platform

Other

Additional Context

Python code. I am using subscription

View original on GitHub ↗

This issue has 5 comments on GitHub. Read the full discussion on GitHub ↗