Behind the Build: Starting Over, Again (and Again)
Let's be ironic for a bit...
Hi friends,
This past week, I finally sat down to start on the CLI for my context utility project!
Remember when I said Working with AI shouldn’t feel like starting over ?
Well I’m not there yet at all!
If you’ve been following along, you know the idea: give AI conversations a little more memory, a little more structure, and a lot less copy-pasting. But getting from “idea” to “working command-line tool” has been… messy.
At least twice, I had to throw out what I started and begin again 🤦♀️
Part of it was technical. I ran into weird import issues, clashing environments, commands that silently failed, and tests that passed when they shouldn’t have. I kept thinking: Why isn’t this working? It’s just a basic CLI.
But part of it was about clarity. Not just in the code, but in my own thinking 😑
What should the first commands be?
How should I structure the folders?
What does a "context" should look like at the filesystem level?
And even more frustrating: how do I ask for help from an AI about a tool I’m building for managing context, when that same AI has no persistent memory of the project?
I’d describe a bug, get a suggestion, try it, break something else, then realize we weren’t even on the same page about the setup. Of course we weren’t, there was no shared context yet. Not even a file to point to. No trace of what I’d already tried. Just me, my notes, and a model doing its best with a partial picture 😩
The meta-irony of all this
I’m building a tool to carry context forward, while stumbling on the very problem I’m trying to solve.
Still, I made progress!
I ended up rebuilding the CLI around two foundational commands:
context init: to create a new project folder with a structured layout for context types
context delete-project: to safely remove a project (with a --force flag, because I’ve learned better, and the hard way)
I added a handful of test, most of which passed, one of which helpfully failed when it should have. That failure forced me to confront a blind spot in how errors were handled, and in how the tool communicated failure to the user (and to the test suite). It was a good reminder that even small utilities need to be explicit, predictable, and easy to reason about.
But more than anything, this session taught me that context isn’t just about storing information.
It’s about shaping how we think, how we debug, how we ask for help.
It’s about who we are and who we want that AI in front of us to be.
Without it, even a simple project can feel like pushing uphill with a blindfold on.
So yes, I got frustrated.
Yes, I started over.
And yes, I’m still excited.
Because every time I hit a wall, I’m reminded that this project, is worth building. Not because it’s fancy, but because it chips away at something real: the silent cost of starting from scratch, again and again, when we work with AI.
What’s next?
In the next session, I’ll start implementing context file creation, loading active context into memory, and wiring up the first summary updates 🎉
Thanks for following along,
Adeline