TIL - Internal thoughts vs external prose

Sometimes the real gems are in the thinking

I was debugging why an old command syntax kept showing up in Claude sessions for tx, it’s been there since v1 and kept appearing. tx msg was a command I originally used to send an agent a message, before moving to the immutable central log.

For some reason, it made sense that Claude would hallucinate that command as some sort of natural conclusion, but today, I tried to track it down. I had rebuilt all the prompts, the CLAUDE.md, there’s no freaking reason it should be showing up again.

I’ve taken to turning on Verbose mode so I can see Thinking, and this time, it proved fruitful. Claude thought:

  Wait, so tx msg IS listed in the mesh-builder skill... but looking at the actual CLI structure mentioned in the core context, I see:

Wonderful! Finally, it was found! However, in the response output, was just apologies and feeble attempts to justify the erroneous inclusion. This was not a suitable approach, as I don’t need apologies, I need information for action.

Given that context, I pulled this prompt language out and it’s going in the system prompt.

When you catch an error in thinking, don't apologize - report the discrepancy with file paths and specifics so the user can fix the root cause.

Time will tell if it bears fruit.

Update: When it comes to specific skills which are for one repo, I’ve taken to having the skill reference the documentation, so the information lives in one place.

Update 2: So far so good, I have not caught this happening again.