I’ve talked about earlier than that I received recognized with consideration deficit dysfunction “primarily inattentive” about 15 years in the past. I don’t take any treatment for it although. I did however it didn’t actually appear to make issues higher. I kind of think about a excessive functioning grownup with pretty vital ADHD-PI although.
There’s a number of issues related to it which have all the time had actual ramifications for empirical analysis, most notably the eye to element and reminiscence. And after I received to Baylor in 2007, I made a considerable funding in constructing an empirical workflow that helped me transfer ahead with analysis whereas concurrently making coding errors occasions that had been random and will get caught, often in time. A variety of it was variations of constant naming conventions of recordsdata and folders, a properly organized folder construction that I used repeatedly for each challenge, automated manufacturing of tables and figures, and a pipeline of code, moderately than one off scripts. I didn’t use GitHub and model management early on, however in time would attempt to shoe horn model management into my workflow as properly. And it labored properly.
The factor about Claude code for analysis is that actually all of these issues it does. It creates replicable code, it organizes the listing properly, its code robotically produces figures and tables, it makes use of constant naming conventions, it’s organized. And since git operations are accessible, model management is simple too.
And so that you’d assume — and I feel for months I did assume this — that subsequently Claude code was merely the newest manifestation of an evolution in my very own course of in direction of larger precision and fewer errors. That’s, the error fee was falling as a result of Claude did these issues properly, so significantly better even than me.
However that is changing into extra obvious to me on the newest initiatives. I’ll name it drift. There’s a sort of drift that occurs for me that I’ve but to totally wrap my head round, and can’t subsequently but design an answer. Listed below are a few of the examples of drift.
-
I’ll study that the script was by no means written. It was advert hoc, created throughout our dialogue, however it didn’t write it down
-
The analytical pattern is shifting over time, implying {that a} script is both being modified unknowingly, or all the basis is coming off its base boards.
-
Remedy items could also be subtly altering from the place that they had been months earlier.
What’s new for me is realizing I’m not catching this in actual time. And the truth that the errors are rising for me however exterior of the conventional location of errors that I’m accustomed to. One thing concerning the analysis workflows pure manufacturing of error has shifted, and it’s in such a means that my regular human capital isn’t conscious of.
I haven’t fastened it but. I’ve been redoing a present challenge from the bottom up. I’m centered on this challenge, not all the workflow, in different phrases. However I do know it’s the workflow. I simply lack the ideas to assist me describe the transition happening, and till I can develop a coherent psychological mannequin — actually in my very own phrases, one thing perhaps visible — I don’t assume my fixes can be true fixes. I feel they are going to be band-aids.
I’m skeptical that the answer is in a /ability to be completely sincere. I feel the issue is linked to my adhd, and abilities actually don’t deal with adhd. I would like extra like a easy system, boiled all the way down to 5-6 main coverage suggestions which can be habitually adopted, and that when habitually adopted, make errors knowable and controllable.
The opposite a part of the issue I’m fairly certain is my present type of utilizing Claude code. When individuals see me use it, they usually comment that my prompts are conversational, very lengthy and over written. It’s how I work. I discuss with and to Claude code. I don’t dictate duties, nor do I am going into plan mode and design the multi-step orchestration. Moderately I’m co-authoring the analysis course of in actual time utilizing dialogue.
One thing about all of that is actually feasting on my reminiscence struggles huge time. I solely discover it as a result of I’ve a brand new model of previous adhd issues with reminiscence. Traditionally, I’d compensate for adhd associated forgetfulness with repetition. I’d do the identical factor repeatedly deliberately through hyper-focus which might create memorization, to offset forgetfulness and poor recall. Hyper-focus is one thing I faucet into incessantly and it helps me memorize the structure of the challenge whereas being productive too.
However now I’m hyper specializing in dialogue and discovery. However not hyper specializing in scaffolding. And this appears to be a supply of the drift. Perhaps not the supply, however it appears downstream from it. It should be as a result of if I put a challenge down for a number of days — even simply two days — I come again unusually amnesiac. And since I didn’t produce the scripts, and there are actually so a lot of them, and so many displays, a few of which seem to have grow to be stale (however which of them?), I don’t have memorization to compensate.
A few of this does appear conceptually like a cousin of the previous issues. Take the abundance of scripts. That’s actually a brand new model of an previous downside I had in grad college the place I’d make 10 variations of recordsdata, often simply individually named by date. The poor man’s model management you would possibly say.
Effectively, that’s type of like what’s occurring right here however it’s not as straightforward to acknowledge bc it doesn’t seem like a multitude. It’s simply it’s a mess — it’s a Claude mess, not a human mess, which is what makes it tougher to acknowledge, establish, diagnose and remedy since his messes to me don’t actually seem like messes. They do look a bit sprawling however it’s like a properly designed type of sprawl.
So I’m persevering with to work on my harness. I haven’t pushed it but, however I added one thing new to it final night time. I took my “lovely deck” idea and constructed a script into the harness that every day would learn the progress logs, which I’ve Claude write after each main session concludes, and writes a stupendous deck written in html accessible from inside the aspect panel of the dashboard.
The gorgeous deck inside my dashboard workflow is my effort at addressing my amnesia. A variety of the progress I’ve made with adhd is to cease pretending I don’t adhd. When you settle for it, you construct programs endogenous to it, not programs that assume it isn’t there. And the attractive deck contained in the dashboard which reads my progress logs to reorient me to the challenge since my final go to is my means of doing that. It’s flippable, it makes use of “lovely” rules (ie lovely figures, lovely tables, lovely slides), and its sole function is to assist the amnesiac bear in mind.
I’m constructing a home from the bottom up. That’s the harness; the harness is just like the bodily expression and seize of my analysis work move. I’m shut, however not there but.
I additionally assume I’ve to modify to a big analog pocket book if I’m going to work this fashion, however for what I can’t fairly articulate it. I must someway map what Claude is doing. As a result of he’s received an excessive amount of latitude — that’s what I’m studying from others, however even from itself — and if I’m going to proceed with my dialogue method, I’ve to determine a system that accommodates it whereas making certain zero error.
