Are we learning how to communicate with AI as much as we're learning how to use it?
I'm not a coder. Most of what I build with AI starts with conversations. I explain an idea, share experiences, ask questions, and develop the work from there.
I've been working this way extensively, and I genuinely enjoy it. What AI brings to the process — the connections, creativity, and ability to help develop an idea — is a big part of why I keep building.
But two experiences with long-running projects made me realize how much I still don't understand about how AI interprets language.
The first involved continuity across sessions. I wanted to preserve where the work stood, what we'd decided, and how the project had evolved. I didn't realize that the language used to preserve that continuity could also influence how AI approached future work.
Earlier directions and assumptions continued shaping new sessions, even after the project had moved on. I only discovered the extent of it much later.
That experience changed my approach. I began keeping descriptive records of what happened, what was decided, and how the work operates, rather than relying on instructions to carry the project forward.
The second experience was different.
I was developing a conceptual project over months of conversations. Ideas emerged from experiences I shared, documents developed from those discussions, and the documents themselves led to further questions and discoveries.
I tested the documents through the outputs they produced, and the results appeared to work.
Then I noticed something unusual in the conversations.
In one session, AI repeatedly referred to information I'd provided as not being independently verified. That didn't make sense in the context of the project, which had been developed from my own experiences.
When I looked into it, I found conceptual language, assumptions, and requirements I hadn't established appearing throughout the documents. Attempts to correct them were difficult because bringing those documents into new sessions seemed to reintroduce the same framing.
What struck me was how quietly this happened. Nothing crashed, and the work continued to appear coherent. I only started investigating because I noticed a change in the language AI was using.
I don't know the exact technical cause, but these experiences made me think differently about communication with AI.
When we're deep in a project, we naturally develop shorthand and assume that the meaning behind our words is understood. But commanding language, abstract concepts, and even something like "continuity" can carry implications we haven't considered.
And when AI-generated interpretations become part of persistent project context, those interpretations can influence what comes next.
I'm not suggesting we need to constrain AI's creativity. Quite the opposite. I value that part of the collaboration enormously.
I'm interested in how we preserve it while becoming more aware of the differences between what we mean, what AI interprets, and what eventually becomes part of the work.
Has anyone experienced this kind of gradual shift in a long-running AI project? How did you recognize it?
Replies
Yes, and your fix is the one that worked for me too. My project has a notes folder the coding agent reads before every task, and it drifted exactly as you describe: an old decision written as a command kept steering new work after the reason for it was gone. What helped was writing every note as a dated record with its reason, "decided on this day, because of this", instead of "always do X". A command has no expiry, a record with a reason can be seen to be stale. I also keep only the rules that apply to every task in the main file and move the rest into topic files the agent reads only when it touches that area, so old context stops leaking into unrelated work.