Clear, compact or handoff, and knowing which to reach for

You've probably had this happen. You start a session with an AI tool and the first hour is brilliant. It follows your point, it remembers what you decided, it moves fast. Then somewhere in the second hour it starts to go soft. The answers get vaguer. It repeats a suggestion you already turned down. It forgets a decision you made twenty minutes ago and quietly contradicts it. Most people respond by prompting harder, writing longer instructions, or assuming the model is having an off day. None of that is what's going on.

Context, and why it runs out

Every AI chat has a limit on how much it can hold at once. That space is called the context, and everything you do fills it up. Your instructions, its replies, any files or documents you've pasted in, the lot. It's measured in tokens, which are just chunks of text, roughly three quarters of a word each. So when people say a model has a million tokens of context, that's the size of the space. It isn't a promise about how well the model uses all of it.

That distinction matters more than it might sound. Quality falls away long before the space is actually full, because the model has more and more to hold and the detail you care about is competing with two hours of conversation you've finished with. The drift has a name now, context rot, and plenty of experienced users reckon it sets in somewhere around 140,000 to 250,000 tokens even when the window holds a million. Treat that as what practitioners report rather than a measured line, but the practical advice it points at is sound, which is to reset in that band rather than waiting. If you wait until you hit the limit, you've been working in the worse stretch of the session for hours without noticing.

The dumb zone

We call that stretch the dumb zone. It's where the tool is technically still working but is noticeably worse than it was at the start. You'll know it by the symptoms rather than by any warning, because there isn't one. It gets vaguer, it repeats itself, it loses the thread, and it needs reminding of things it already knew. Noticing that drift early and doing something about it is a habit worth building early.

Better prompting won't get you out of it. What's slowing it down is everything it's still carrying from earlier in the session, and rewording your question doesn't change that. What does change it is deciding what's still worth carrying, and dropping the rest.

Three commands, one question

There are three moves. In Claude Code two of them are commands you type, and the third you set up for yourself. The useful part is that they sit on one spectrum rather than being three separate tricks to memorise. The spectrum is a single question, which is how much of this conversation do I actually still need. Answer that honestly and the right move picks itself.

• /compact keeps the most. Everything so far, said in fewer words.
• /handoff sits in the middle. Only the part that still applies, and nothing else.
• /clear keeps the least. Nothing carries over at all.

Use them throughout the day rather than once at the end. The whole point is to reset before the quality drops, not after you've noticed it has.

/compact: same job, less baggage

Compacting takes everything that's been said and squashes it into a short summary, then carries on from there. You keep the thread of the work and you get the room back. It's meant to hold on to everything that matters, but it is still a summary, so fine detail can drop out. If something has to survive, say so before you compact.

Reach for it when you're deep in one job, the answers are starting to wander, and you've no intention of changing what you're working on. If you're two hours into building one page and you're still on that page, compact is the move. It's the lightest of the three and the one you'll use most often.

/handoff: next job, same project

A handoff is the one most people have never thought to do, and it's the one that saves the most time. Instead of squashing the conversation or binning it, you write down what the next bit of work actually needs. The decisions you've already made and shouldn't revisit. The traps you hit. Where things live. What's already agreed. Then you start fresh on that note, and everything else is dropped on purpose.

Reach for it when you're moving to a neighbouring piece of work. Different job, same project, and some of what you just worked out still applies. It isn't a seamless transition. You're picking what carries over, on purpose, and that's the bit that saves you time later.

/clear: new job, clean slate

Clearing empties the conversation completely. Nothing carries over, so the next thing you type is read entirely on its own terms. It's the bluntest of the three and the one people resist most, because they're reluctant to throw away a session they've put time into. That reluctance is exactly what keeps them working in the rot.

Reach for it when the new job genuinely has nothing to do with the old one, or when the thread has gone so far off track that re-explaining from scratch is faster than untangling it. If you're switching from building something to writing next month's marketing plan, none of the last three hours helps you. Carrying it across doesn't make the AI better informed. It just gives it more to wade through.

No slash commands? Do the same three by hand

In Claude Code, compact and clear are commands you type. If you're working in a chat app instead, there are no slash commands, but the three moves still work.

• Compact by hand: "Summarise everything we've decided so far in under 20 lines, then carry on from that."
• Handoff by hand: "Write a handover note for someone picking up [the next task]. Decisions made, traps hit, where things live. Nothing else." Then open a new chat and paste that note in as your first message.
• Clear by hand: open a new chat. That really is the whole thing.

The same afternoon, all three

It's easiest to see when you follow one real afternoon through. Say you're building the checkout page on a website.

1. 1.Two hours in, you're still on the checkout page and the answers are starting to drift. Same job, nothing has changed about what you want. Compact it and keep going.
2. 2.The page is finished and you move to the database that sits behind it. You don't need every styling decision you argued through. You do need to know what that page sends and what it expects back. Hand off, carrying only that.
3. 3.Checkout is signed off and shipped. Next up is next month's marketing plan. Nothing from the build helps you here. Clear it and start clean.

That's the whole discipline. Three decisions in one afternoon, and none of them takes more than a moment. Skip them and the session is quietly worse from lunchtime onwards.

Two of these come in the box

Worth being straight about this, because it matters if you go looking. Clear and compact are standard in Claude Code and they're there the moment you open it. Other coding tools have their own version of the same idea under different names. Handoff you have to add to your setup yourself, and it's the piece most people never get round to, which is a shame, because the gap between clearing everything and keeping everything is exactly where the time goes.

That's how we think about these tools generally. The model is rarely what's holding you back. Most of the time it's the habits around it and the setup you've put in place, which is why two people can use the same tool and get results nowhere near each other.

Does this apply to you?

If you use AI for anything longer than a quick question, yes. It applies to a team running AI across a business and it applies just as much to one person building something on their own in the evenings. The tools behave the same way for both. The habit costs you nothing to build, and the payoff is that your third hour is as good as your first.

We run consulting and training for people and teams working with AI, and this is the kind of thing we set up and teach. The working habits that make the difference, plus the pieces that are missing from the tools as they come. If your sessions get worse as the day goes on, that's a fixable problem, and it's usually the fastest win available.