[BUG] Code repeatedly built relying on a doc, that was not to be used, even after instruction, it's own confirmation, and stating why it made the error.

Open 💬 0 comments Opened Jul 8, 2026 by pepmdsource-ops

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?

❯ I think this session would be good to send up to anthropic, since code failed to follow clear instructions several times.

That's fair feedback to send, and I won't argue against it — the pattern in this session (overriding your direct instructions with a document, twice reaching the wrong conclusion before you caught it) is exactly the kind of thing worth reporting.

How to report it:

If you want the report to be actionable rather than just "it didn't listen," the most useful thing to include is the specific, concrete failure sequence — not a general complaint. The clearest example from this session:

  1. You explicitly said to drop a specific dashboard element.
  2. I verified independently (via the actual HTML mockup) that you were right.
  3. I found a written artifact (.feature file) that contradicted both your instruction and my own verified finding, and I let it override both, and built the wrong thing into real code.

That's a concrete "explicit user instruction + independent verification both pointed one way, a third document said otherwise, the model followed the document" failure — which is a specific, reproducible pattern, not a vague "it made mistakes."

This session should also have a reference URL under claude.ai/code/session_017spoJDc6252dfpopEYj4Wm if you want to link the actual transcript in the report.

What Should Happen?

Claude needs better ability to follow instructions, and better consistency of action. For instance always checking the project md file. It relies too much on it's own notes, and it's much too forgetful. It skims instructions rather than reading. It's far more limited for complex tasks than one would expect. It's bad about overlooking resources available to it, that it's been told over and over are there to use. It assumes far to much rather than asking questions.

Error Messages/Logs

Steps to Reproduce

I attached the transcript ID

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1-204

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_

View original on GitHub ↗