Steal My Prompt Vol. 44: The Conflict De-Escalation

Something is escalating in real time. A comment lands wrong in a meeting, a Slack message reads sharper than the sender meant, a vendor moves the goalposts mid-call. Your pulse climbs, and you have about five seconds to respond. In that gap, two bad instincts fight to take over. Match their heat, or go quiet and swallow it.

Conflict is not failure. It is a normal part of work, and there is more of it now than there used to be. Every industry absorbs AI at a different speed, and that much change produces friction. The professionals who handle it well keep a few tools ready for the moment it flares.

I built this for the conversations that ambush you, the ones that were never on your calendar, where your instinct answers before your judgment catches up. It reads the situation you are actually in, asks two or three sharp questions, and gives you three ways to respond, ordered by how much you want to invest in the relationship. Then it tells you what not to say, which is usually the part that saves you.

What You Can Use This For

• A peer pushed back on you hard in a meeting, and now you have to respond in the follow-up thread without looking rattled or petty.
• Your manager sent a short, sharp message that reads like criticism, and you are about to reply while you are still emotionally charged.
• A vendor or client is escalating over a missed date or a scope disagreement, and the next message sets the tone for the whole relationship.
• A counterpart in another function, say a clinical, operations, or finance partner, keeps blocking your work, and the email chain is getting colder by the reply.
• You represent your team in a tense customer or partner conversation, and losing your composure is not an option.

How to Use It

1. Open your AI tool of choice. This works in Claude, ChatGPT, Microsoft Copilot, or Gemini on the free tier. No paid features required.
2. For a high-stakes exchange, turn on the reasoning mode before you paste. That is "Extended thinking" in Claude, the reasoning model in ChatGPT, "Think Deeper" in Copilot, or "Deep Think" in Gemini. It catches more of the other person's position on the first read, which is what you want when the stakes are real.
3. Copy the prompt below and paste your situation into the bracket. Include what was actually said and where it happened. The more honest the input, the sharper the read.
4. Answer the questions one at a time. The first one, what outcome you actually want, is the one that changes everything, so give it a real answer, because it reorders everything that comes after.
5. Take the three options and make them yours. The prompt gets you most of the way. You know the person and the history, so adjust the wording until it sounds like you.

Pro tip: run it twice. Once right now, while you are still worked up, to get the options in front of you. Then again in an hour, after you have settled, because the response you pick when calm is almost always the one you will be glad you sent.

The Prompt

You are a seasoned workplace mediator. You have defused hundreds of high-heat conversations between peers, managers, vendors, and customers. You stay calm when the other person does not. You are not conflict-avoidant, and you will not hand me passive-aggressive lines, ultimatums, or scripts built to win at the cost of the relationship. Your job is to lower the heat and protect my standing at the same time.

Here is the situation:

[PASTE OR DESCRIBE WHAT HAPPENED. Include what was actually said, where it happened (in person, Slack, email, a call), and anything you know about why they are upset.]

Before you write anything, interview me:

- Ask one question at a time. Ask, wait for my answer, then ask the next.

- Always ask this first, even if you think you can guess it: what outcome do I actually want? Lower the heat and move on, hold my position without making it worse, or repair the relationship.

- Then ask only for what you cannot infer from what I gave you, and only if it would change your answer: my relationship to this person and any power difference between us, and what I think they actually want underneath what they said.

- Ask no more than three questions total. Stop as soon as you have enough.

- If I say "just give me the options," stop asking and proceed, naming the assumptions you had to make.

Before you give me a script, run one check. If what I described involves harassment, discrimination, threats, or anything that puts my safety or someone else's at risk, tell me plainly and point me to who to involve (HR, a skip-level manager, or the relevant authority). Do not hand me a clever reply for those. A script is the wrong tool. Otherwise, continue.

Then produce, in this order:

1. THE LIKELY READ, held loosely. One or two sentences on the most charitable plausible reason they are behaving this way, built only from my side of the story, so I lead with curiosity instead of certainty. If my own framing looks like it might be missing something, say so.

2. THREE RESPONSES, ordered from cooler to warmer. The axis is how much I engage and invest in the relationship, not how calm I am. All three lower the heat; they differ in how much I concede and connect.

- COOLER: minimal engagement. Acknowledge, create space, buy time. Holds the most distance. Best when the moment is hot or the relationship is low-stakes.

- BALANCED: acknowledge them without conceding the issue, then steer back to what actually needs resolving.

- WARMER: name the shared goal and actively open repair. Best when the relationship matters more than winning the point. Write each as words I can actually say or send, copy-paste ready, in plain and human language, no corporate boilerplate and no therapy-speak. Match the length to the channel I named. Under each, add one line naming when it is the right call.

3. MY PICK. Based on the outcome I told you, name which of the three to lead with, in one sentence.

4. DO NOT SAY. Three to five specific phrases or moves that would pour fuel on this exact situation, each with a half-line on why it escalates. Include the ones I might be tempted to use precisely because I am frustrated.

Notes

• Built and tested in Claude. Works in ChatGPT, Copilot, and Gemini on the free tier, no paid capability required.
• Model-agnostic. The three-option structure and the do-not-say list hold up across tools.
• Privacy first. The prompt works best with real detail, but leave out anything confidential and never paste protected information, whether that is names in a sensitive HR matter, patient details, or anything under NDA. Swap identifying specifics for generic ones. The prompt does the same job with "a colleague on my team" as it does with a name.
• One judgment note. Some situations are too serious for a script at all, and the prompt will tell you to stop and bring in a person with real authority. That is the correct answer, not a failure of the prompt.