You Should See a Doctor, a Claude Doctor

A recent publication from Anthropic came out on how to work with the Claude 5 models. It discusses how they've reduced Claude Code's system prompt and one piece of advice they give is to run /doctor and see what, if anything, you should change when working with these models.

The /doctor command in Claude Code acts as a health check and diagnostic tool that automatically inspects your installation, configuration, permissions, and context settings to identify and fix setup issues.

Given my entire setup is built around memory and context that remain harness agnostic, I was curious what I would find and if I could improve my experience with Claude 5 models (because I'm not loving Opus 5).

In the end, I saved about 3,700 tokens per session from the cleanup. On a million-token context window that is negligible. However, the information it surfaced that has nothing to do with tokens garnered important results.

reading my own context is actually an audit

First there was the breakdown of skills I have. I have a lot of skills, and in fact, fifteen of them had never once been invoked. I picked them up somewhere along the way thinking "oh that's cool", and then clearly had no need for them.

Next, my CLAUDE.md file told Claude to read another file called bot-briefing.md at the start of every conversation. My bot-briefing.md file is a memory index plus a small extras file, concatenated by a script. It exists so that I can remain harness agnostic. However, when it comes to Claude Code, memory is injected automatically and doesn't need this bot-briefing.md file. So what is happening is that I end up loading my entire memory index, then Claude also reads a file that contains my entire memory index. It's doing this every single session. That's about 2,500 tokens of pure duplication, in a file I built specifically to solve a context problem.

There's a lot of juggling happening with all these AI products and approaches. It is easy to overlook when things are getting messy. Running this audit is another proof point on why maintenance of tools, yes even AI tools, is necessary.

This report also surfaced bugs. Back in June, I fixed a bug in shared code across health agents that I run for me, my husband, and my dad. Two of these agents run exactly the same. My dad's agent is set up just different enough that it's a couple of extra steps to get fixes deployed. For whatever reason, the deploy didn't fire and it wasn't reported back to me that the fix was not implemented. For about five weeks there's been a bug silently tracking quantities incorrectly. So that's finally fixed.

Then this audit found a few other discrepancies on projects and cron jobs I had been working on. I'm not surprised while also I had no idea. AI is letting us build faster than ever before. For those of us with all these ideas in our head, AI has been the key to unlocking our creativity. It also is fragmenting our attention. Or at least my attention. I'll have multiple agents running multiple projects and experiments at one time. It's multi-tasking on steroids. Except something is going to break. A person can only focus on so many things before something is missed. Running an audit is a good way to find what's been dropped.

The other takeaway was discovering entries in my "active projects" that Claude can access. This had grown to around twenty-five entries and turns out eleven of them are automations that run fine on a schedule and don't need to be top of mind for me or Claude at the start of every conversation. It's just creating noise.

Have I learned how to better work with Claude 5 models? Nope. In fact, I'm finding Claude 5 Opus challenging. However, I have fixed several bugs and cleaned up a system that was clearly in need of it. I guess it's a reminder that yes, every so often, you should visit the /doctor.