Rehearse the hard conversation with AI before you have it

It's 10pm. You have a 1:1 with your manager tomorrow and there's one thing you need to say out loud — the thing you've been avoiding for three weeks.

You've written the Slack DM four times. The cursor sits after the greeting. You read your own draft back and it sounds fine, which is not the same as sounding fine to them.

Two options form in your head: send it cold tonight and deal with the reply in the morning, or wing it tomorrow and hope the right words appear. Both of them are a first attempt in the live meeting.

Why the first attempt keeps happening in the live meeting

Writing a draft is not the same as rehearsing it. When you reread your own words, the voice in your head is polite, charitable, and already agrees with you.

The other person is none of those things. They have a different memory of last quarter, a different sense of what the decision means, and their own pattern for reacting when someone raises a concern.

Rereading is a closed loop. Rehearsal is a loop that pushes back.

What AI practice is actually good for

An AI chatbot will never be your manager. It doesn't know the last three conversations you had with them, or which exact sentence from 2024 still makes them flinch.

What it is good for is the thing most rehearsal lacks: a partner that talks back, instantly, as many times as you want, without getting tired of you.

The evidence for this is not speculation. A 16-week controlled study of Grammarly use in undergraduate writers found measurable gains in clarity and structure — feedback that would have taken a human tutor hundreds of hours to deliver. A 2025 trial with pre-service teachers found AI-assisted speaking practice significantly improved oral performance and reduced public-speaking anxiety. The pronunciation coach ELSA reports that roughly 90% of users feel their speech improves and 95% feel more confident after sustained practice.

The common thread isn't that AI is a great coach. It's that a mediocre partner who shows up on demand beats a good partner who shows up never.

A prompt shape that works

The mistake people make is asking AI to write the message for them. That shortcuts the rep — the message gets sent, but you still haven't rehearsed anything.

Four prompts do most of the work.

You are my manager. Here's your context:
- You pushed this deadline through last week.
- You reacted to my last pushback by going quiet, then overriding it.
- You care about looking in control in front of your director.

I'm going to open a conversation. Reply in character — short, realistic. Don't soften. React the way the above says you'd react.

Here's the first sentence I plan to say:

"I want to flag something about the Q3 rollout before we lock it in."

Give me three ways this could land badly, and why.

Rewrite this message at three directness levels — soft, direct, blunt — without changing the underlying content.

[paste your draft]

For each version, tell me what it signals about the sender.

Here's my draft. What am I not saying that probably matters? What am I hinting at instead of naming?

Situations

Feedback conversation

You've been sitting on one piece of feedback for your report for two weeks. Rehearse the opening and the first pushback.

You are Sam, my direct report. I'm about to give you feedback about the Tuesday draft landing late. You tend to react defensively at first, then get reflective 30 seconds in. React in character. Keep replies under two sentences.

Deadline pushback

You are my manager. You've just asked for feature X by Friday. The reality is it needs A + B cut to hit Friday, or the full scope in 2 weeks. Roleplay you pushing back when I surface the trade-off. Don't let me off the hook easily.

Disagreeing with your manager

You are my manager. In our last meeting you said we'll ship X this way. I think there's a risk you missed. Roleplay how you'd respond if I raised it — including the version where you feel like I'm questioning your judgement.

Where AI falls short

The limits are real and worth naming, so you don't over-trust the output.

• Relational history. AI doesn't know that you and your manager already had "that fight" in February. Its model of "your manager" is generic.
• Cultural nuance. Research on AI translation warns that nuance and idiom flatten out. The same goes for tone advice — AI produces something that reads right to the average reader, which is not the same as right to this specific person.
• The script trap. Rehearse the same phrase 12 times and you'll deliver it with a slight recitation flavour. Audiences feel it. Practise the shape of what you're saying, not the exact words.

If it still feels fake

After each AI exchange, close the tab, stand up, and say your next reply out loud in your own voice.

Your voice will compress, simplify, and drop the stiff bits. That's the version you'll actually use tomorrow.

Try this today (90 seconds)

Open a chat with your AI of choice. Paste this:

You are [the person]. You tend to react to [thing] by [reaction]. I'm about to say the following to you:

"[your opening sentence]"

Reply in character. Short. Don't soften.

Then:

• Read their reply.
• Say your real response out loud — don't type it.
• Do it twice more with their pushback escalating.

That's the whole rehearsal. Ninety seconds, three reps, one opening line that now has some weight on it.