Use an LLM as an Independent Judge in Difficult Conversations.

TL;DR

• Giving and taking feedback is hard, and it is hardest exactly when you need it most: when things are moving fast and a Slack thread starts to feel tense
• The hack: copy the entire thread into ChatGPT, Claude, or whatever you use, and ask an open question - “What are your thoughts when reading this conversation?”
• You get a surprisingly useful outside read on two things: is this actually escalating or does it only feel that way, and where does the other person genuinely have a point
• Then bring that evaluation into your next 1:1 and let the model explain its reasoning. It moves the discussion from “you vs. me” to “us vs. the text”
• Ask openly, never leading. The moment you ask “am I right here?”, you get a cheerleader instead of a judge
• For anything confidential, run a local model. Qwen 3.8 on your own laptop is a genuinely good judge for this task, and nothing ever leaves the machine
• Caveats apply: confidentiality, missing context, and never use the output as a weapon

The problem: you are the worst possible judge of your own conflict

Feedback is hard in both directions. Giving it well is hard. Receiving it without flinching is harder. And judging a conversation you are inside of is close to impossible.

When a thread starts to feel tense, your brain does two unhelpful things at once. It inflates the temperature - a terse one-liner from your boss at 22:00 reads like an ultimatum when it was probably just someone typing on a phone. And it protects your position - you replay the argument in your head and, funnily enough, you win every single time.

This gets worse under speed. Difficult conversations do not happen when things are calm. They happen during a launch, a re-org, a missed deadline, an incident. Exactly when nobody has the patience to be diplomatic and everybody is reading messages between meetings.

The classic advice is “sleep on it” or “ask a trusted peer”. Sleeping on it helps a bit. Asking a peer works, but peers are expensive: they have their own opinion about the other person, their own stake in the outcome, and you are pulling them into office politics whether you meant to or not. Also, at 23:00 on a Thursday, they are not available.

The hack: make the LLM the independent judge

Here is what I do instead, and it has helped me multiple times.

Copy the full thread. Paste it into ChatGPT or your favourite LLM. Ask an open question.

Something like:

What are your thoughts when reading this conversation?

That is it. No framing. No “my boss is being unreasonable, right?”. No context about who is who beyond what the thread already contains.

What comes back is usually some version of: here is what each person seems to want, here is where the tone shifted, here is where the misunderstanding started, here is a point person B makes that person A never actually answered.

And that last part is the valuable one. The model does not care about your ego. It reads the text, not the history. It has no stake in the outcome, no memory of the last three times you disagreed with this person, no interest in being invited to lunch tomorrow.

Two things typically come out of it:

1. Is this really heating up? Often the answer is no. The model reads a thread you experienced as a confrontation as a normal, slightly terse work exchange. That alone is worth the two minutes - it stops you from writing the reply that would actually have started the conflict.
2. Where is the other person right? This is the uncomfortable one. Nine times out of ten there is a point in there that you skipped over because you were busy defending yourself. Seeing it written out neutrally is much easier to accept than hearing it from the person you are arguing with.

The open question is the whole trick

The single most important detail: ask openly, never leading.

“What are your thoughts when reading this conversation?” gives you a judge.

“Don’t you think my boss is being unfair here?” gives you a lawyer. LLMs are agreeable by design and will happily build you a beautiful case for whatever you hinted you wanted to hear. You will feel great and learn nothing.

A few prompts that work well, roughly in order of increasing bravery:

• “What are your thoughts when reading this conversation?”
• “Summarise what each person in this thread actually wants.”
• “Where in this thread did the tone change, and why?”
• “Which points made by the other person were never addressed?”
• “Steelman the other person’s position.”
• “If you had to name one thing I am getting wrong here, what would it be?”

If you want to be really honest with yourself, anonymise the names before pasting, so the model cannot infer who is “the boss” and who is “the report”. Hierarchy leaks into the judgement otherwise.

Then bring it into the 1:1

This is the part that turns a private sanity check into an actual tool.

In your next 1:1, put the evaluation on the table. Not as “the AI says you are wrong”, which would be the worst possible use of this. As: “I was not sure how to read our thread from Tuesday, so I had a model read it cold. Here is what it said. Does that match how you experienced it?”

Two things happen:

• It depersonalises the conflict. You are no longer two people defending positions. You are two people looking at a third, boring, neutral description of what happened. The discussion moves from “you vs. me” to “us vs. the text”.
• It stays grounded. Difficult conversations drift into feelings, history, and vague impressions within about ninety seconds. A written third-party read keeps pulling you back to what was actually said.

And you can let the model explain itself. If the other person disagrees with the judgement, ask it right there in the meeting: why do you read it that way? Often the explanation is more useful than the verdict, because it makes the misunderstanding visible - “the request in message 4 was phrased as a question, so it was easy to read as optional”. That is the kind of thing neither of you would have articulated on your own.

For confidential threads: use a local model

There is an obvious objection to all of this, and it is a good one: the conversations where this hack helps most are exactly the ones you should think twice about pasting into a cloud service. Performance topics, a conflict with your boss, a disagreement about someone’s role, customer names, numbers that are not public yet.

The answer is simple: run the judge on your own machine.

This is not a compromise any more. A local model on a decent laptop is entirely good enough for this job - and in my experience Qwen 3.8 is a really good judge. Reading a conversation, working out what each side wants, spotting where the tone shifted, and naming the point that never got answered is not a task that needs a frontier model. It needs careful reading, and a local model does that well.

The practical setup is unspectacular: LM Studio or Ollama, download the model, paste the thread, ask the same open question. I wrote up the details - hardware, model choices, and the two settings that will bite you - in Running LLMs Locally.

Two things make this more than a privacy workaround:

• Nothing leaves your laptop. No terms of service, no retention policy, no “was this used for training?” question, no company policy to check. You can paste the thread verbatim, names included, and stop worrying about it.
• You will actually be honest. This one surprised me. When the model runs locally, you paste the real thread instead of a sanitised version, and you ask the question you actually want answered. A judge you have quietly censored is not much of a judge.

My rule of thumb: cloud model for a normal work disagreement where the stakes are tone and timing, local model the moment a name, a number, or somebody’s job is in the text.

Why this works better than doing it yourself

It is not that the model is smarter than you about human relationships. It is that it has three properties you cannot have about your own conflict:

• No stake. It does not need to be right, does not need to be liked, and will not have to work with either of you tomorrow.
• No history. It reads this thread, not the last two years.
• No cost to ask. You can ask a genuinely embarrassing question - “am I the difficult one here?” - at midnight, with nobody watching. That is a real advantage. Most people never ask that question out loud.

It is also just a lot less work than the alternative, which is either stewing about it or scheduling a coffee with a colleague to relitigate the whole thing.

Caveats - and they matter

This is a hack, not a management framework. A few honest limits:

Confidentiality first. A work conversation can contain names, salaries, customer details, or performance topics. With a cloud model you are handing all of that to a third party, so use whatever your company has actually approved and anonymise where you can. And for anything genuinely sensitive - health, performance ratings, someone’s job - do not compromise, just run it locally. That is what the section above is for.

The model only sees the text. It does not know that this person has been under pressure for a month, that there is a history, or that “sure, let’s do that” from this particular colleague means “over my dead body”. Its read is a data point, not a verdict.

It will flatter you if you let it. See above. Leading questions get you a mirror. Open questions get you a judge.

Never use it as a weapon. “ChatGPT agrees with me” is the fastest way to make a tense conversation genuinely toxic. The output is a conversation starter you both examine, not evidence you present.

It does not replace HR, a therapist, or a real decision. If a situation is serious - harassment, a performance issue, a legal question - this is not the tool. Go to a human.

Takeaway

Feedback is hard, and the harder it gets, the worse we are at judging it. Handing the raw thread to a model and asking an open question is a two-minute move that gives you something rare: a neutral read on whether the situation is really on fire, and where the other person actually has a point.

Then take that read into the 1:1 and look at it together. It keeps the discussion grounded, it takes the personal heat out, and more often than not it ends the disagreement faster than another round of Slack messages ever would.

And when the thread is confidential - which is often exactly when you need the second opinion most - run it through a local model instead. Qwen 3.8 on your own laptop does the job, and the question of what you are allowed to paste simply disappears.

Cheap, fast, and it has saved me more than once.

More

• Related: Running LLMs Locally on getting a good model onto your own machine
• Related: Three Strike Dismissal in One-On-Ones on trusting your gut in 1:1s
• Photo on top by krakenimages