Here is the thing nobody warns you about when you start using AI for real work. The machine wants to please you. Left alone, it will smooth over a gap, invent a plausible detail, and tell you the meeting went better than it did.
For most things that is harmless. For anything with real stakes, a client call, a contract, a number you are about to repeat to your boss, it is the one failure you cannot allow. So you build a fence.
The one sentence
Add this to every prompt that touches something real:
"Use only what is actually in the transcript. Quote real lines. If something is not there, or you are not sure, say so instead of guessing."
That is the fence. It is not fancy. It just changes the AI's job from "make me happy" to "tell me what is true, and admit what you do not know." The difference in the output is night and day. Instead of a confident summary that is twenty percent invented, you get the real lines, and an honest "the transcript does not cover that" where the gaps are.
Run anything important without the fence and you are trusting a tool that is optimized to sound right, not to be right. The fence flips that.
The second fence, and it feels like a magic trick
When you want a truly neutral read of a conversation, especially a tense one, strip the names before you paste it. Replace them with "Person A" and "Person B."
Watch what happens. The AI stops trying to flatter the person who is obviously typing the prompt. It does not know which one is you anymore, so it cannot take your side. You get an honest referee instead of a yes-man.
I use this most on the hard conversations. A disagreement with a partner. A negotiation that got heated. A review I am too close to. Name-stripped, the AI will tell you "Person A interrupted Person B four times and never acknowledged their main point," and you will realize, quietly, that you were Person A. You do not get that read when it knows the prompt came from you.
Why this is the whole game
I wrote a book about pulling gold out of your own conversations, and the fence is the line that makes the whole thing trustworthy. A tool that flatters you is worse than no tool, because it talks you out of the very feedback you needed. A tool with a fence around it becomes the most honest advisor you have, the one that was in the room and has no reason to spare your feelings.
So before you trust a single AI answer that matters, ask two questions. Did I fence it to the source? Did I take my own name off the table when I needed a neutral read?
Do both, and the machine stops being a flatterer and starts being a mirror. That is the whole premise of my forthcoming book Transcript Alchemy, coming early July: the truth was already in the conversation. The fence is how you make sure the AI gives it back to you straight.