How do you use Claude to prepare for a hard conversation at work?

It's Thursday night. Tomorrow at 10 you have a 1:1 where you need to bring up the PR that sat in review for four days, again. You open Claude, type "help me tell a coworker they're slowing the team down," and get back three tidy paragraphs that sound like an HR memo. You close the tab. Tomorrow you say nothing.

The assistant wasn't the problem. The brief was. An AI chat is a good place to think before a hard conversation, but only if you treat it like a sharp colleague who doesn't know the situation yet, and not like a script vending machine.

Brief it like a colleague, not a search box

The assistant only knows what you tell it, and it will fill every gap with the most generic version of your situation. Give it the shape of what happened, not your verdict about it.

What to give it:

• What you saw. Specific events, with rough dates. "The last three PRs sat in review for three to five days" beats "they're slow."
• How often. Once is a different conversation from the fifth time. The history of the pattern decides how heavy the conversation should be, not how annoyed you are.
• What you've already said. Including nothing. This changes the advice more than anything else.
• Your relationship. Peer, your manager, someone on your team, a partner on another team. Say what authority you have and don't have.
• What you want. A specific change, not "for them to get it."
• Why it's hard for you. "They're senior to me." "We're friends." "Last time I raised something, they went quiet for a week." This is the part people skip, and it's usually the real reason the conversation hasn't happened.

What to leave out:

• Full names. Use a role or an initial.
• Health, family or personal details you learned in confidence.
• HR case files, compensation, legal matters.
• Confidential company or customer data.
• Long pasted private threads. Summarize the message instead.

Check your company's policy on AI tools before you start; many have one. A good rule: if you wouldn't want the other person to read what you pasted, rewrite it until you would.

If you do this often, put the standing context (your role, your team, how you like to talk) in your custom instructions or a project, so each new chat starts from what you've already told it.

Prompts that work at each stage

The Accountability Dial has five stages: Mention, Invitation, Conversation, Boundary, Limit. They're time-separated, and most situations resolve at the first two. Knowing which stage you're at keeps you from walking in with a Boundary when a Mention would have done it.

Stage	What to ask the assistant	What good output looks like
Mention	"Help me name this in one sentence, today, without making it a big deal."	One line, specific, no lecture. Often praise.
Invitation	"It happened again. Help me ask what's going on without assuming."	A genuine question, not a disguised accusation.
Conversation	"Help me describe the impact on the work, not my frustration."	Concrete costs: the handoff, the client, the team's week.
Boundary	"Help me state what needs to change and by when, once, plainly."	A clear ask with a date, no threats, no hedging.
Limit	"Help me say this is the end of the conversation on this issue, and what happens next."	Calm, short, consistent with what was said before.

Start every session by asking which stage you're actually at:

Here's the pattern and what I've already said about it. Based on the history, not on how frustrated I am, which stage of the Accountability Dial does this sound like? If you think I'm skipping a stage, say so.

For a Mention:

I want to mention this today in one or two sentences, in Slack or in person. It's the first time I'm raising it. Give me three versions: one plain, one warm, one that starts with something they did well this week. No more than 30 words each.

For an Invitation:

I mentioned this two weeks ago and it happened again on Tuesday. Help me write a question that invites them to look at what's underneath it. Don't assume a reason. Flag anything in my wording that sounds like I've already decided.

For a Conversation or Boundary:

Here's what I need to change and by when. Write it as I'd say it in a 1:1: short sentences, no corporate phrasing. Then tell me which sentence I'm most likely to soften in the moment.

How to get pushback instead of flattery

Chat assistants lean toward agreeing with you. You describe the coworker as unreliable, and the assistant helps you tell an unreliable coworker they're unreliable. That's the trap.

Ask for the disagreement before you ask for the words:

Before you help me, argue the other person's side as well as you can. What might they say happened? What am I not seeing because I'm inside this? Rate from 1 to 5 how sure you are that my read is fair.

Then try the cold read. Open a new chat, give only the facts with no adjectives, and ask what it thinks is going on. If its answer differs from your story, that difference is the most useful thing you'll get all night.

One more:

Am I raising this because it matters to the work, or because I'm annoyed? Ask me two questions to find out.

You may not like the answers. That's how you know they're working.

Rehearse it before you say it

Ask the assistant to play the other person, and give it a realistic reaction to aim for:

Play my teammate. You're a bit defensive and you'll point out that the reviews were slow because priorities changed. Respond to my opening line, and stay in character for four or five exchanges. Then step out and tell me where I got vague or started over-explaining.

Run it twice: once where they get defensive, once where they agree quickly. The quick agreement is the one people aren't ready for. They nod, and you keep talking because you'd prepared for a fight.

Then say your opening sentence out loud. Not in your head. If it sounds like a memo when you hear it, it is one.

Turn it into a first sentence you'd actually say

The opening is the only part worth preparing word for word. After that, you listen.

Here's my draft opening. Cut it to one sentence I could say across a table. Keep my words where you can. Remove anything that sounds rehearsed, apologetic, or like a verdict.

What a good result sounds like:

"Hey, I want to bring up the review times on the last few PRs. Got a few minutes after standup?"

"Can I name something from the client call? The update went out after the call instead of before, and I want to figure out how we avoid that."

Short, specific, no apology, and it leaves room for them to talk. If the assistant hands you three sentences, ask for one.

Mistakes to avoid

• Asking for a script. You'll sound like you're reading, and the other person will respond to the script instead of you.
• Letting it pick your stage. Severity isn't the same as history. A first-time miss on a big project is still a Mention.
• Pasting everything. More private detail rarely makes the advice better. It just widens the blast radius if you've misjudged what's safe to share.
• Taking the first draft. The first output is the average answer. Push back on it at least once.
• Preparing for an hour, then waiting a week. Prep is meant to make the conversation easier to start. If it becomes a reason to delay, close the chat and send the one sentence.
• Forgetting the follow-up. Ask the assistant for a two-line recap you can send after, so the agreement lives somewhere other than both of your memories.

For what to do once you're in the room, see how to have a difficult conversation at work.

Doing this with Ren inside the assistant you already use

If you'd rather not rebuild the brief every time, Ren works inside Claude, ChatGPT, Cursor and Microsoft 365 Copilot through Ren's MCP server. You paste Ren's MCP URL into the assistant's connector settings and sign in with your Ren account; in Microsoft 365 Copilot, an admin adds Ren once in Copilot Studio. Setup is at tryren.com/mcp.

The assistant asks before it uses Ren, Ren asks before reading your calendar or email, and anything Ren drafts stays a draft until you approve the exact words. You can add Ren's skills, such as 1:1 Prep and Close-Out, Dial-Aware Feedback, and Daily Brief, and in Claude type / to use them. Chats can be opened back in the Ren app.

On the flattery problem specifically: in an independent blind study, Ren scored 4.22 out of 5 on pushback and ownership, compared with 2.84 for ChatGPT, 2.68 for Copilot and 2.48 for Gemini. The full results are in The Pushback Report.

---

Related: What is the Accountability Dial? · How to have a difficult conversation at work · Privacy-first AI at work, explained · The Pushback Report