The Meeting Coach
Build an AI that grades your meetings and tells you how to fix them

Nobody tells you that you talk too much

Most professional feedback stops arriving somewhere around your first management job. People stop telling you that you talk too much, that you open with a disclaimer instead of a claim, that you asked one question in fifty minutes. Not because you fixed it. Because it got awkward to say.

A transcript doesn't find it awkward.

This came out of our bonus episode, where Ceci described the meeting coach she built for herself. Her framing was blunt:

"It's probably like a side effect of my insecurity." Cecilia Manduca — EP 11

Then she described what it does, and Bea's reaction on tape was "This is brilliant. Another one that I will copy."

Three things, and you almost certainly have all of them

You do not need: a paid notetaker, a specific company, connectors, an ops team, or permission from anyone. If you can get a transcript of a meeting you were in, you can run this today.

Do not run this on the cheap tier

The single most common mistake with this workflow is running it on a cheap, fast model. Transcripts are long. A one-hour meeting is 8,000 to 12,000 words. Small models lose the middle of long documents, and the middle is where the useful evidence lives.

As of August 2026:

Why Luna and Haiku are the wrong call here: both are built for cheap, high-volume, short tasks. On long-context recall — which is exactly what reading a full transcript is — Luna scores around 41% against Sol's 91%. It will confidently quote things nobody said. For a workflow whose entire value is accurate evidence, that's disqualifying.

Get one transcript

Pick a meeting from the last week where you did most of the talking. Not your best meeting. That's the point.

Export the transcript as plain text. Almost every notetaker has a copy-transcript or download-transcript option. You want the raw dialogue with speaker labels, like this:

Alex (00:04)
So I wanted to walk through the Q3 numbers before we get into the roadmap.

Priya (00:11)
Sure. Do you want me to pull the dashboard up?

Speaker labels matter — without them the coach can't tell your talking from theirs. If your tool doesn't label speakers, most will let you tag them once and remember.

Run the coach

This is the whole workflow in one prompt. Paste it into Claude or ChatGPT, fill in the bracketed parts, paste your transcript underneath.

It works standalone, in a plain chat window, with no setup at all. Step 4 turns it into something reusable.

You are an executive meeting coach. You are honest, specific and unsentimental.
You are not encouraging. Your job is to make me better, not to make me feel good.

Below is the transcript of a meeting I was in.

MY NAME IN THE TRANSCRIPT: [your name exactly as it appears]
MY ROLE: [e.g. Chief of Staff / Head of Finance / Founder]
WHAT THIS MEETING WAS FOR: [one line — e.g. "get the team to commit to a
  launch date" or "understand why churn moved"]
WHAT A GOOD OUTCOME WOULD HAVE BEEN: [one line]

Analyse ONLY what is in the transcript. Follow these rules absolutely:

1. Every single claim you make must be supported by a VERBATIM QUOTE from the
   transcript. If you cannot quote it, do not claim it.
2. Never invent, paraphrase or reconstruct a quote. Quote exactly or say
   "no evidence in transcript".
3. If the meeting was genuinely good, say so. Do not manufacture problems to
   seem useful. An honest "three things went well and one thing to watch" is a
   valid result.
4. Do not comment on my personality. Comment on behaviour I could change on
   Monday.

Produce exactly this, and nothing else:

## THE NUMBERS
- My share of the words spoken: X% (and the rough split for everyone else)
- Questions I asked: X total — of those, X were genuinely open (invited
  thinking) and X were closed or rhetorical. List the open ones verbatim.
- My longest uninterrupted turn: approximately X words, on the topic of [topic].
  Quote the first line of it.
- Times I interrupted someone mid-sentence: X. Quote each one.
- Times someone tried to speak and I continued: X. Quote each one.
- Decisions actually reached: list them. For each, say whether it has a named
  owner and a date, or does not.
- Open loops left at the end: anything raised and never resolved.

## WHAT WORKED
Two or three things I did well, each with the quote that proves it.

## THE VERDICT
Three sentences maximum. The single most important pattern in how I showed up.
Lead with the claim. No preamble, no cushioning.

## THREE FIXES
Exactly three, ordered by how much they'd change the outcome. For each:
- **The behaviour:** what I did, with the verbatim quote
- **Why it cost me:** tied to the stated purpose of this meeting
- **Say this instead:** an actual rewrite of my actual words. Not general
  advice. The specific sentence I should have said.

## ONE THING TO TRY IN MY NEXT MEETING
A single, concrete, testable action. Something I can either do or not do, and
know afterwards which it was.

TRANSCRIPT:
---
[paste transcript here]
---

What comes back

You'll get numbers you can't argue with and a verdict you probably can. Both are useful. The numbers are the part that changes behaviour — most people are genuinely surprised by their own talk share, and nobody guesses their longest monologue correctly.

Rule 3 matters more than it looks. An AI asked for criticism will always find criticism. Explicitly permitting "this was fine" is what makes the bad reports trustworthy.

Make it mean, or don't

Ceci's version is deliberately harsh:

"It's really mean sometimes, but I want it mean." Cecilia Manduca — EP 11

That's a preference, not a requirement. Add one of these lines to the prompt:

Be blunt to the point of uncomfortable. Do not soften anything. Assume I am
experienced and can take it. Skip all praise unless it is genuinely notable.

Be direct but constructive. Lead with what worked. Frame fixes as experiments
rather than failures.

Turn it into something reusable

Pasting a prompt works, but you won't keep doing it. Here's how to install it properly on each platform.

On Claude

Option A — a Project (easiest, 5 minutes)

1. In Claude, create a new Project called Meeting Coach.
2. Open the Project's custom instructions and paste in the prompt from Step 2 — everything from "You are an executive meeting coach" down to but not including TRANSCRIPT:.
3. Replace the bracketed fields with your real name and role once, permanently.
4. From then on, every new chat in that Project only needs the transcript pasted in. Everything else is remembered.

Option B — a Skill (better, reusable, invokable by name)

1. Create a folder called meeting-coach.
2. Inside it create a file called SKILL.md.
3. Put this at the top of the file:

---
name: meeting-coach
description: Analyse a meeting transcript and score how the user showed up —
  talk share, questions asked, longest monologue, interruptions, decisions
  reached. Returns a verdict and three specific fixes with rewrites. Use when
  the user shares a meeting transcript, asks how a meeting went, asks for
  feedback on their own performance in a call, or says "coach this meeting".
---

4. Underneath the ---, paste the full prompt from Step 2.
5. Zip the folder so you get meeting-coach.zip, then rename it meeting-coach.skill.
6. Upload it in Claude's skill settings.

Now you can say "run meeting-coach on this" and paste a transcript, in any conversation.

Option C — automate it (the version Ceci actually runs)

If you have Claude connected to your calendar and your notetaker, you can go further and have it run without you asking. Ask Claude, in a Project that already has the skill:

Set this up as a scheduled task. Every weekday at 18:00, look at my calendar
for meetings that happened today where I was a participant. For each one that
has a transcript available in [your notetaker], run the meeting-coach analysis.
Then give me ONE combined summary for the day: the numbers per meeting, and a
single pattern you noticed across all of them. Do not send me one report per
meeting — I will not read five reports.

That last sentence is the difference between a habit and an ignored notification.

On ChatGPT

Option A — a Project

1. Create a new Project called Meeting Coach.
2. Put the Step 2 prompt into the Project instructions, with your name and role filled in.
3. Paste transcripts into chats inside the Project.

Option B — a custom GPT

1. Go to Explore GPTs → Create.
2. Name it Meeting Coach.
3. In Instructions, paste the Step 2 prompt.
4. Under Conversation starters, add: Here's a transcript, coach me.
5. Turn off web browsing and image generation. It doesn't need them, and switching them off stops it wandering off to look up meeting-facilitation blogs instead of reading your transcript.
6. Save as private.

Option C — Tasks

ChatGPT's scheduled Tasks can prompt you rather than fetch for you. Set one for Friday afternoon: "Ask me to paste this week's toughest meeting transcript, then run the Meeting Coach analysis." Less automatic, but it survives having no connectors at all.

Add the tracking layer

This is the part missing from most versions of this workflow, including the one on the episode. A verdict tells you about one meeting. A trend tells you whether you're actually getting better.

After each run, append one line to a running file or sheet:

Then once a month:

Here is my meeting log for the last month. Do not coach individual meetings.
Tell me: what is measurably improving, what is not moving at all, and which
fix I have now been given three times and clearly haven't implemented.
Quote the dates.

The third one is the useful one. Being told the same thing repeatedly, with dates attached, is harder to ignore than being told it once.

Make it fail loudly, and audit it

Two rules from the episode that apply to everything you build, not just this.

Make it fail loudly

If you automate this and the transcript isn't available, the run must tell you. A workflow that silently does nothing is worse than no workflow, because you'll believe it's running. Add to any scheduled version:

If you cannot find a transcript for a meeting, do not skip it silently. Say
explicitly: "No transcript found for [meeting name] — coach did not run."

"If they can't reach the information, make it say: Hi, I'm the fairy, I couldn't reach the information." Beatrice Aliprandi — EP 11

Silence is not success.

Audit your own setup

Once you've built two or three of these, point your strongest model at all of them and ask it to find what's broken. On the episode, Bea did exactly this and it found an authentication flaw in one of her own scheduled tasks, left over from reconnecting some connectors.

Go through every skill, scheduled task and automation I have set up. For each
one, check: what data can it read, what can it write or send, whose credentials
is it using, and what happens if a connection breaks or a permission changes.
Flag anything where the access is broader than the job requires. Be specific
about the risk, not general.

Run it on Claude Fable 5 or ChatGPT Sol. This is the one job where you want the expensive model — a reviewer that rubber-stamps everything is worse than no reviewer.

When not to do this

Honest limits, because a guide that only sells you the upside isn't a guide.

• Sensitive meetings. Some conversations should not be transcribed at all, and some shouldn't leave your machine. Ceci retired a different automation partly for this reason — the output would have sat on a calendar she'd then have to make private. If a meeting is confidential, it's outside this workflow. Full stop.
• Interviews and performance conversations. Don't point this at meetings where you're assessing another person. The transcript then contains their material, not yours, and coaching yourself on it slides quickly into analysing them.
• Meetings where talking a lot was the job. A talk share of 80% in a presentation is correct. Tell the coach the purpose in the prompt or it will scold you for doing your job properly. This is why the WHAT THIS MEETING WAS FOR field exists.
• The first week will be unpleasant, and that's fine. Everyone's first three reports are worse than they expected. Judge the workflow on week four, not day one.

When it misbehaves

A meeting coach isn't really about meetings

It's the smallest possible version of a bigger idea: the parts of your job you can't see are the parts you can't improve, and a transcript is just a mirror you can't argue with.

The busywork was never the job. But neither was flying blind.