Claude Code 54: Amnesia

I’ve mentioned before that I got diagnosed with attention deficit disorder “primarily inattentive” about 15 years ago. I don’t take any medication for it though. I did but it didn’t really seem to make things better. I more or less consider a high functioning adult with fairly significant ADHD-PI though.

There’s a few things associated with it that have always had real ramifications for empirical research, most notably the attention to detail and memory. And after I got to Baylor in 2007, I made a substantial investment in building an empirical workflow that helped me move forward with research while simultaneously making coding errors events that were random and could get caught, usually in time. A lot of it was versions of consistent naming conventions of files and folders, a well organized folder structure that I used repeatedly for every project, automated production of tables and figures, and a pipeline of code, rather than one off scripts. I did not use GitHub and version control early on, but in time would try to shoe horn version control into my workflow as well. And it worked well.

The thing about Claude code for research is that literally all of those things it does. It creates replicable code, it organizes the directory well, its code automatically produces figures and tables, it uses consistent naming conventions, it’s organized. And since git operations are accessible, version control is straightforward too.

And so you’d think — and I think for months I did think this — that therefore Claude code was simply the latest manifestation of an evolution in my own process towards greater precision and fewer errors. That is, the error rate was falling because Claude did those things well, so much better even than me.

But this is becoming more apparent to me on the latest projects. I’ll call it drift. There is a type of drift that happens for me that I have yet to fully wrap my head around, and cannot therefore yet design a solution. Here are some of the examples of drift.

1. I will learn that the script was never written. It was ad hoc, created during our dialogue, but it did not write it down

2. The analytical sample is shifting over time, implying that a script is either being modified unknowingly, or the entire foundation is coming off its base boards.

3. Treatment units may be subtly changing from where they had been months earlier.

What’s new for me is realizing I’m not catching this in real time. And the fact that the errors are emerging for me but outside of the normal location of errors that I’m accustomed to. Something about the research workflows natural production of error has shifted, and it’s in such a way that my normal human capital isn’t aware of.

I haven’t fixed it yet. I’ve been redoing a current project from the ground up. I’m focused on this project, not the entire workflow, in other words. But I know it’s the workflow. I just lack the concepts to help me describe the transition taking place, and until I can develop a coherent mental model — honestly in my own words, something maybe visual — I don’t think my fixes will be true fixes. I think they will be band-aids.

I am skeptical that the solution is in a /skill to be perfectly honest. I think the problem is connected to my adhd, and skills really don’t address adhd. I need more like a simple system, boiled down to 5-6 major policy recommendations that are habitually followed, and that when habitually followed, make errors knowable and controllable.

The other part of the problem I am pretty sure is my current style of using Claude code. When people see me use it, they often remark that my prompts are conversational, very long and over written. It’s how I work. I talk with and to Claude code. I don’t dictate tasks, nor do I go into plan mode and design the multi-step orchestration. Rather I am co-authoring the research process in real time using dialogue.

Something about all of this is really feasting on my memory struggles big time. I only notice it because I have a new version of old adhd problems with memory. Historically, I would compensate for adhd related forgetfulness with repetition. I’d do the same thing repeatedly intentionally via hyper-focus which would create memorization, to offset forgetfulness and poor recall. Hyper-focus is something I tap into frequently and it helps me memorize the layout of the project while being productive too.

But now I am hyper focusing on dialogue and discovery. But not hyper focusing on scaffolding. And this seems to be a source of the drift. Maybe not the source, but it seems downstream from it. It must be because if I put a project down for a few days — even just two days — I come back unusually amnesiac. And since I didn’t produce the scripts, and there are now so many of them, and so many exhibits, some of which appear to have become stale (but which ones?), I don’t have memorization to compensate.

Some of this does seem conceptually like a cousin of the old problems. Take the abundance of scripts. That is literally a new version of an old problem I had in grad school where I’d make 10 versions of files, usually just separately named by date. The poor man’s version control you might say.

Well, that is kind of like what’s happening here but it’s not as easy to recognize bc it doesn’t look like a mess. It’s just it is a mess — it’s a Claude mess, not a human mess, which is what makes it harder to recognize, identify, diagnose and solve since his messes to me don’t really look like messes. They do look a bit sprawling but it’s like a well designed kind of sprawl.

So I am continuing to work on my harness. I haven’t pushed it yet, but I added something new to it last night. I took my “beautiful deck” concept and built a script into the harness that each day would read the progress logs, which I have Claude write after every major session concludes, and writes a beautiful deck written in html accessible from within the side panel of the dashboard.

The beautiful deck within my dashboard workflow is my effort at addressing my amnesia. A lot of the progress I have made with adhd is to stop pretending I don’t adhd. Once you accept it, you build systems endogenous to it, not systems that assume it isn’t there. And the beautiful deck inside the dashboard which reads my progress logs to reorient me to the project since my last visit is my way of doing that. It’s flippable, it uses “beautiful” principles (ie beautiful figures, beautiful tables, beautiful slides), and its sole purpose is to help the amnesiac remember.

I am building a house from the ground up. That’s the harness; the harness is like the physical expression and capture of my research work flow. I am close, but not there yet.

I also think I have to switch to a large analog notebook if I’m going to work this way, but for what I can’t quite articulate it. I need to somehow map what Claude is doing. Because he’s got too much latitude — that’s what I’m reading from others, but even from itself — and if I’m going to continue with my dialogue approach, I have to figure out a system that accommodates it while ensuring zero error.