Tell it what changed.

If AI keeps using old details after you change your plan, it's because you added the new version without retiring the old one. Everything you said earlier is still sitting in the conversation. The model doesn't know which parts you threw out. It just sees two versions and tries to honor both.

Prompt Fluency is communicating so the model doesn't have to guess. Here's a guess you might not notice you're asking for: which version is the real one right now?

Picture this. You run a small bakery. Early in the chat you say:

"Wedding order for Saturday. 150 guests, three tier vanilla cake, plus 200 cupcakes. Build me a bake schedule and an ingredient list."

Good answer. Then the bride calls. You type:

"Okay, she wants lemon now, and it's 120 guests."

You get a new schedule. Looks fine. But the cupcakes are still 200, the ingredient list still has vanilla bean on it, and the bake times are still built for the bigger cake. You read it and think the model wasn't paying attention.

It was paying attention. To everything. You told it lemon and 120. You never told it the cupcakes got cut, or that vanilla was gone for good, or that the whole schedule should start over from the new numbers. So it patched the parts you mentioned and kept the rest.

Same thing shows up everywhere.

A youth soccer coach moves the tournament from Saturday to Sunday, and the snack signup still says Saturday.

A landscaper drops the retaining wall from a bid, and the labor hours still include it.

A landlord changes the renewal from 12 months to 6, and the letter still mentions the yearly discount.

Each time, the new detail made it in. The old detail just never left.

So say what changed, and say what's dead. Plain English.

"Update on the wedding order. Changes: lemon instead of vanilla, 120 guests instead of 150, cupcakes are cut completely. Nothing else from before carries over unless I say so. Rebuild the bake schedule and the ingredient list from scratch with these numbers."

"Tournament moved to Sunday. Saturday is off entirely. Redo the snack signup and the parent email so Saturday doesn't show up anywhere."

"The retaining wall is out of the bid. Remove its materials and its labor hours, then show me the new total and what dropped."

"Lease renewal is 6 months now, not 12. The yearly discount doesn't apply anymore. Rewrite the letter."

Look at what changed. You stopped tossing a new fact on the pile and hoping. You told the model what the current truth is and what to throw away. The answers come back clean, because it isn't trying to keep two plans alive at once.

Think of it like a whiteboard in a shop. If the job changes and you just write the new number next to the old one, the next guy who walks in has to guess which one counts. If you erase the old one and write "CHANGED" next to the new one, nobody guesses.

No shame here. I do this too. I'll be ten messages deep, change my mind, and type the new idea like the model already knows I dropped the old one. It doesn't. It only has the words. If I didn't say the old thing is gone, it's still on the board.

And if the chat has gotten long and messy, sometimes the cleanest move is one fresh message that states the whole current plan in one place. Not a patch. The plan.

A quick check when your plan shifts:

Ways to tell it what changed

• Start with "Changes:" and list them
• Say what's gone, not just what's new
• Say "nothing else carries over" when you mean it
• Ask it to rebuild from the new numbers instead of patching
• Ask it to list what it removed so you can check
• When it's a mess, restate the whole current plan in one message

When an answer has one stale detail hiding in it, ask yourself: did I say the old thing was dead? Odds are you only said what was new.

You're not repeating yourself. You're removing the need to guess. That's Applied Intelligence Mastery in practice: say what changed, say what's gone, and let the model work from one plan instead of two.