Every one of these tools has a limit on how much it can hold in its head at once. Not how much it knows. How much of this particular conversation it can keep in front of it while you work.

I think of it as runway. You're not out of fuel. You're running out of room to land safely, which is a different problem and it has a different answer: land on purpose, before the end.

That one idea is behind most of the long projects I've finished with AI over the last couple of years. It's also the difference between a build that gets better each week and one that quietly gets worse until you abandon it and blame the tool.

At the start of anything big, I say some version of this:

That's it. That's the whole trick.

What it buys you is permission to work loose. If you're worried about running out, you ration. You skip the back and forth, you don't explore the second idea, you try to nail everything in one pass. Hand the watching to the machine and you can argue with it for an hour, chase something that doesn't pan out, and come back, because somebody's minding the gauge.

Then at the right moment it stops on its own, writes the state down, and hands you a paragraph to start over with. You paste it into a fresh window and keep going with a full tank.

This works for anything that runs longer than a sitting. Software you're building. A training program. A website. A process you're documenting. Anything where week three depends on what you decided in week one.

Here's where I did it wrong for a while, and it's the part worth stealing.

I used to write those handoff notes for the machine. Machine-to-machine plumbing. State, decisions, next steps, go.

Then a kickoff prompt sat unused for three days because the week got away from me, and when I finally opened it I had no idea what I'd been doing. The AI would have been fine. I was the one lost.

So now I say it out loud when I ask for one: future me is going to read this too, and future me won't remember any of this.

That single line changes what comes back. It stops being a status dump and becomes a note to yourself that the AI happens to be able to read as well. What we decided and why. What we already tried that didn't work, so you don't burn a day repeating it. What's next. What not to touch.

I run a lot of things at once and they blur together. That note is what lets me pick up a build on Tuesday that I put down on Friday without spending an hour reconstructing my own thinking. The value isn't that the AI remembers. It's that I do.

You'd think the failure mode here is handing over too little. Mine has always been the opposite.

I'd stuff everything into the handoff. Every reference document, all the background, the whole history, on the logic that more can't hurt. And the note designed to protect the runway would eat the runway. Fresh conversation, already half full, before we'd done a thing.

The fix is to point instead of carry. Keep a few plain files where the detail lives, and let the handoff be short and say the decisions are in this file, the standards are in that one, start here. It pulls what it needs when it needs it. The note goes back to being a paragraph.

Here's the one concrete tell, and I've never seen anybody else name it.

When you see a message about the conversation being compacted or summarized to keep going, you already overshot. That's not a warning. That's the overrun area at the end of the runway. It's still working, but it's now working from a summary of your work instead of your work, and things start falling out that you won't notice until much later.

Set it up right and you never see that message, because you land before it. If you do see it, don't push on. Stop there, take the handoff, start clean.

This is a good habit, not a rule, and I'd say so to anybody on my crew. Nobody needs a procedure for it. A short project doesn't need any of this.

And read the handoff before you use it. It's a summary written by something that's low on room, which is exactly when a detail is most likely to go missing. Thirty seconds of checking beats rebuilding a decision you made two weeks ago.

On your next long project, open with one sentence: watch your runway, and when it gets low, stop and write me a kickoff prompt for a fresh start — written so I understand it too, not just you.

Then let yourself work loose, because somebody's watching the gauge. And tell me the longest thing you've ever tried to build this way and where it came apart — I'd put money on it being right around the point everything started feeling vague.

Wednesday: the other half of this, and the stranger half — asking the AI to tell you what you did wrong.