Your AI tool doesn't get dumber over time. It gets buried.
I run a Claude Project for juvaii's client work: case studies, proposal templates, an active-engagements tracker, my own working instructions, one new chat per project. It used to answer a one-line message instantly. Then it started taking two and a half minutes to reply to something as simple as "yes."

So what happened? Isn't AI supposed to just work? Does AI need maintenance? To answer that, you need to actually look at your model's reasoning trace. Yes, you, non-technical person working with AI. That's becoming the skill every professional needs to develop right now, not prompting better.

Here's what was happening: on almost every message, Claude was rereading a long log file and rewriting the whole thing, whether or not anything I'd said actually required a change. My first instinct was "too much information in there." Wrong diagnosis. The real problem was that I had six different places holding information, and I'd never decided what belonged in which one, or which one Claude should actually trust. In other words, a disaster.

So instead of blaming the model or give more instructions, I inspected the workflow itself: what context it was retrieving, what it was trying to persist, which source it considered authoritative and what actual tool operations it had to perform.

Turned out there were two separate problems hiding as one:
- Context architecture: working notes, durable memory, source documents, operational data and history were all mixed together, so Claude kept redoing work nobody asked for.
- Tool architecture: a simple spreadsheet update was being handled as read everything, rebuild everything, replace everything, because the tool didn't expose the one small operation the task actually needed. (It's connected to Google Drive).

Now, every time I ask my Project to do its job, it's fast again and we work together like the colleagues we are. But Claude still needs a manager. ME 🙂.

That's the same fix behind almost every slow, unpredictable AI setup I evaluate for clients: nobody decided which source is authoritative, so the tool tries to keep everything synced everywhere, all the time. Slower, not smarter. The problem was never the model. It was ARCHITECTURE.

If an AI tool you rely on has quietly stopped feeling worth the wait, check what it's being asked to remember and rebuild before you blame the model. Happy to compare notes if you're untangling something similar.

PS: the one piece I never automated, updating my own tracker by hand, is still the fastest part of the whole system. Sometimes the "no tool" step is the right architecture too.