Why does an AI chat's context fill up so fast?

I’m building a workflow where an AI handles work alongside me.

Every time an AI responds, there’s a cap on how much material it can draw on at once, usually called its context. In this workflow, the task background the system auto-loads eats into that same space.

What needs doing, how far along it is, who picks it up next: all of it lives in one task list.

The idea was that every new chat could skip the re-explaining. The AI would pick up where things stood straight from the task records, knowing what came next and what to watch for.

So I made sure the AI could pick up any task cold. Each one gets a full background: why it exists, what constrains it, what to watch for. The system loads every still-open task’s write-up automatically, the moment a new chat starts.

While the list stayed short, that was just convenient.

Once the list grew to fifty tasks, the problem showed up.

Full background comes at a cost

Say I only meant to touch one task today. What the AI got at the start was the write-up for all fifty.

Every other task’s background rode along too, whether or not it had anything to do with this one.

That made “starting fresh” its own kind of contradiction. I’d clear the old chat to make room for the next thing, and the system would immediately move a whole pile of material back in. Before any work began, the background material had already taken up part of the space.

I’d set the system up to hand over full background every time, to save myself from explaining things twice. What I hadn’t noticed was that I’d been treating “might need this later” as “need this right now.”

See what’s open first, then open the one you’re doing

I still need those detailed write-ups. Delete them and the next handover, mine or the AI’s, has to untangle everything from scratch again.

So what changed this time was how the material gets read, not whether it exists.

A new chat now opens with a short index: what’s on the list, who’s currently on it. Only once a task is picked does its full write-up get opened.

It’s the same as any handover: check the to-do list first, find the item you’re taking over, then go read what goes with it. Everything else stays exactly where it was, findable whenever it’s actually needed.

The work rules the AI reads every turn got the same treatment: an index first, pointing at which file covers which situation, instead of reading every detail up front.

The material’s all still there, the opening just got lighter

Measured against the same task list, before and after, the data auto-loaded at the start dropped by about 93 percent.

That number is about how much text loads at the open, not what the whole session ends up costing. Once the AI actually works a task, it still reads that task’s full detail.

But at least now, starting over doesn’t mean reading the background for every unfinished task first.

This change made me rethink something. Writing the material out clearly is only half a handover. The other half is making sure whoever, or whatever, picks it up next can find the right piece exactly when it’s needed.

I used to care mostly about whether I’d explained everything. Now I ask one more question too: does this need reading right now?