Part 3 of 5

Managing the Context Window

Every conversation has a memory budget — the context window. As it fills, answers get slower and a little worse. The skill isn't avoiding that; it's knowing when to start fresh, and carrying your progress across.

A conversation filling up

35% full

Drag to fill the window with more turns, longer documents, more back-and-forth.

FreshFillingStart fresh

Plenty of room — keep going.

Early in a chat the model has lots of budget and little to re-read. Answers are quick and sharp.

Answer quality
High

Why a full window means worse answers

The model has no memory between turns — every time you hit enter, it re-reads the entire conversation to answer. A long chat means more to re-read each turn, more chances to fixate on something early, and eventually no room left for new material. We unpack this in What Happens When You Ask.

When to start fresh

You've finished one task and are moving to a clearly different one.

It keeps repeating an old mistake or circling back to something you've moved past.

Replies are getting noticeably slower or vaguer than they were at the start.

The tool itself warns you the context is nearly full.

Carry your progress across — the checkpoint

Don't lose your work when you start fresh. Before you open a new chat, ask the model to write a handoff you can paste in. (This is itself a machine-readable twin — see Part 4.)

The checkpoint prompt
We're going to continue this in a fresh conversation. Write me a
handoff note I can paste into the new chat so you'll have full
context. Include:

- The goal we're working toward
- Key decisions we've made and why
- The current state / what's done so far
- The exact next step
- Anything I should paste along with it (files, data, links)

Keep it concise and use Markdown headings.

Watching the meter in your tools

Chat apps hide this, but the more capable agent tools show you exactly how full the window is. Two I use — pricing in Tools I Use.

In Claude Code

The status line shows how much of the window is used. Two commands help:

/contextsee what's filling it up
/compactsummarize & reclaim room without losing the thread

In Claude CoWork

CoWork shows a context indicator on the running task. When it climbs toward full, wrap up the current task and start a new one — pasting in a checkpoint if you're continuing the same work. Scoping each task narrowly keeps the window healthy from the start.