The hardest conversations in your life don't get retries. A salary negotiation, critical feedback to a teammate, a breakup, telling your boss you're quitting — you get one take, live, with a real person whose reaction you can't control.
So most of us rehearse in the shower, say the wrong thing anyway, and then spend weeks replaying it.
I kept running into this problem as a founder, so I built cosskill: an AI communication simulator where you rehearse high-stakes conversations with realistic AI personas before they happen. This post is about what the product does, and what I learned building it — including the parts that surprised me.
What rehearsing actually looks like
You pick a scenario (negotiate a raise, deliver negative feedback, end a relationship, quit gracefully), pick a persona, and talk. You can type or speak. The AI doesn't roleplay a pushover — it pushes back, interrupts, gets defensive, and responds like an actual person would.
Then you get an evaluation of the conversation: what landed, what didn't, and how it moved. You can rewind to any point and try the moment again — including the hardest 30 seconds.
There are 20 personas across operators, strategists, philosophers, and psychologists, plus a mission mode with live meters that track how the conversation is going.
Pricing, plainly: it's free to try — 3 messages without an account, 10 after signing in, with all personas and scenarios available. Pro is $12.90/month or $129/year for unlimited.
Three things building it taught me
1. A roleplay bot that agrees with you is useless
Early versions were too polite. The "other side" folded too easily, and practicing against it trained avoidance, not skill. The single biggest product decision was making personas that genuinely push back. If your rehearsed opener survives a realistic counterparty, it'll survive a real one — or at least you'll know what the counter-move looks like.
2. The rewind is the feature
Live conversations don't have undo. That's exactly why practice should. Being able to redo the worst moment of a conversation — immediately, with a persona that reacts differently each time — is where most of the learning happens. Repetition builds fluency; fluency is what keeps you articulate when you're actually nervous.
3. Feedback has to be concrete
"You did great!" is worthless. The evaluation had to name specific moves: you opened with an apology that undercut your ask; you stated your number once and then kept talking; you de-escalated a personal attack without conceding the point. Concrete diagnosis is what you can act on in the retry.
A 10-minute rehearsal you can do today
- Pick the conversation you've been avoiding.
- Write down your ask in one sentence, with your number/decision stated once.
- Run the scenario with a persona that will push back.
- When it goes sideways, rewind and redo that moment — don't restart the whole thing.
- Read the evaluation, adjust one thing, run it once more.
If you want a deeper framework, we published free guides like How to Have a Difficult Conversation and How to Negotiate Your Salary.
Honest limits
An AI persona is not your manager, your partner, or your teammate. It can't predict how a specific person will react, and it shouldn't be a substitute for real advice on legal, medical, or financial decisions. What it can do is take the edge of panic off — you walk into the real conversation having already heard the pushback, so you're responding instead of improvising.
That's the whole bet: fluency comes from rehearsal, and hard conversations deserve a rehearsal.
If you try it, I'd genuinely like to hear what the AI pushed back with — that's the part we iterate on fastest.
cosskill: https://cosskill.com — free to try, no account needed for the first few messages.
Top comments (0)