Rehearse the question that sends you rambling

The move Before your check-in, type your answer to the question you dread. Have AI check it against your own notes for the one line that would mislead, without rewriting it. Then try again in your own words.

Your weekly check-in is at 10, and you know the update cold. The test import into the sales team's new customer database worked, the full import starts Thursday if IT finishes its setup by Wednesday, and the vendor still hasn't said whether the old call notes can come across.

Then Lena, your manager, asks the one that always gets you: "So are we still on track for November 2?"

That's where your answer wanders. You open with the test import, drop the "if," and somewhere in the middle the vendor "should get back to us this week," a date nobody gave you.

So rehearse that question before 10. Type your answer the way you'd say it, have AI check it against your update notes for the one line that would mislead, then try again.

You walk in with an answer you've already given once, in your own words, with the "if" and the open question still in it.

The usual prep is to ask AI for likely questions and polished answers, then read them over. A good answer you've read feels like one you could give. Lena's question is where you find out whether it is.

Your own answer shows you where it breaks, before the meeting does.

The prompt

Help me rehearse one answer for my weekly project check-in. I'll give you my update notes, the question I expect and my answer, typed the way I'd say it out loud.

Who's in the meeting: [names and roles]
My update notes: [paste them]
The question I expect: [the one that sends you rambling, or leave blank]
My answer, typed the way I'd say it: [type it, or leave blank]

1. If my notes are missing or too thin to check against, ask for them. If the question is blank, ask me one likely follow-up based on my notes. If my answer is blank, stop and wait for it.
2. Check my answer only against my update notes. Find the one thing most likely to leave a wrong impression: a condition I dropped, a date or promise my notes don't support, or an open item I made sound settled. Quote my words and the line in my notes they don't match, and say what's different. Don't guess what anyone will think, repeat or do because of it. If nothing important is off, say so.
3. Don't rewrite my answer or give me a model answer. Ask me to try again in my own words, then wait.
4. Check my next try the same way. Once it holds up, ask me the question someone would most likely ask next, and wait.
5. If my notes can't answer a question, help me practice saying it's still unknown. Don't invent a date or promise for me.
6. Use only these notes, even if earlier messages in this chat say something else. This is practice for what I say, not a prediction of how anyone will react.

Your notes are the referee, so AI checks the answer against what you wrote, not against what sounds confident. One gap per try keeps each fix small enough to make before 10. "Don't rewrite my answer" keeps the second try yours instead of a script you read. If you don't know which question to expect, leave that line blank and the prompt tells AI to ask you one first.

Try it on this made-up example

Who's in the meeting: Lena Ortiz, my manager, and the two sales team leads. Lena passes the switch date up to the sales director.

My update notes:
Project: moving the sales team's contacts from the old database into the new customer database
- Done: Test import of 500 contacts finished Tuesday. 12 records had broken phone numbers; fixed by hand.
- Next: Full import of all 18,400 contacts starts Thursday, if IT finishes setting up the new fields by end of day Wednesday. If that slips, the import moves to next Monday.
- Open: The vendor hasn't said whether five years of old call notes can come across. Asked last Tuesday. No reply yet, and no date promised.
- Target: Sales switches to the new database on Monday, November 2.

The question I expect: "So are we still on track for November 2?"

My answer, typed the way I'd say it: Yeah, I think we're in pretty good shape. The test import went really well, we only had a handful of broken records and we fixed those by hand. The full import starts Thursday, so that's on schedule. The one open thing is the call notes, but the vendor should get back to us this week, so I don't see that holding anything up. So yes, November 2 should be fine.

What Claude flagged

We sent the prompt and that example to Claude Opus 5.5 twice, each time as a single plain-text message with no tools. Both replies went after the vendor line. Here's the second one, with its formatting removed:

The one thing most likely to leave a wrong impression is the call notes.

What you said: "the vendor should get back to us this week, so I don't see that holding anything up."

What your notes say: "The vendor hasn't said whether five years of old call notes can come across. Asked last Tuesday. No reply yet, and no date promised."

What's different:
- "This week" is a timeline your notes don't support. The vendor hasn't committed to any date.
- "I don't see that holding anything up" makes an open item sound settled. Your notes don't say whether the call notes affect November 2. They only say the question is unanswered.

Try the answer again in your own words. I'll check it against your notes the same way.

It names one gap, quotes both sides and hands the answer back to you. The other gap waits for your next try: "The full import starts Thursday" drops the condition about IT's setup.

For comparison, we sent Claude the same example with a plain request: "Help me practice my answer to the question I expect at my weekly project check-in." It caught both gaps. Then it wrote you a revised answer in your speaking style, and that script had you saying "I'll follow up today," a promise your notes never made.

Fix the vendor line, then the Thursday one, and your own answer might end up like this:

November 2 is still the plan. Two things I can't promise yet. The full import starts Thursday only if IT finishes setting up the new fields by Wednesday. If not, it moves to next Monday. And the vendor hasn't told us whether the old call notes can come across. We asked last Tuesday and haven't heard back, so I don't have a date.

The first version could have sent Lena to the sales director with a vendor date nobody promised.

What stays yours

AI checks your answer against your notes, so stale notes mean a stale check. It can't tell you how Lena will react, so the prompt tells it not to guess. It won't hear how you sound, either: this is typed practice, with no voice mode or recording. We didn't test the later steps, a back-and-forth chat, other Claude models or other AI tools, which may follow the no-rewrite rule differently. Whichever you use, make it one your company has approved for project details.