[BUG] Bullshit detector

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

With Claude 4.6, it's 'smarter' but less reliable and certainly not consistent. I notice the 'guessing' has increased to the point where I can't really trust Claude to code, and I have to review or do things myself, and I'm exploring how to use 'dumber' models like Codex in conjunction with Claude for 'thinking.'

I tracked some (not all) of the things that Claude got wrong this week using Claude (and, of course, it lost track of quite a few).

⏺ 62 entries (#58 is missing — gap in the log). Breakdown by category:

Wrong info ~21 Confident incorrect answers
Time wasted ~15 Inefficient approaches, unnecessary loops
Ignored memory ~12 Didn't read MEMORY.md or KB before acting
Broke something ~11 Production deploys, corrupted data, etc.
Repeated mistake ~10 Same failure across multiple sessions
No Autonomous Deferral ~8 Skipped work without being asked to
Deflection ~2 Blamed tooling/environment instead of own error

The dominant pattern is wrong info + ignored memory — giving false confidence about things I didn't actually verify, often because I didn't check what
was already documented. Today's UniFi session contributed to both: three premature "try it now" calls that should be in here, plus the SSH hostname
guessing.

I've built my own 'KB/Memory' systems - and I have a structured approach to developing features that involves Claude not doing things, but writing scripts to do them and then another script to test the result.

...and still.

Claude is good, but being smarter is not better if it comes at the cost of trust/reliability.

What Should Happen?

There should be some sort of 'cost' to Anthropic/Claude for repeated mistakes.

Error Messages/Logs

Steps to Reproduce

Just use the thing.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.71 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_