Guide

The transcript as a mirror: a prompt for honest feedback on your meetings

Paste the transcript of a meeting you ran into Claude or ChatGPT, say which speaker is you, and send this: “Do not summarise. List each question I ducked, where I changed the subject, and who barely spoke. Quote the lines. If none, say so.” No special feature is needed, only an ordinary chat on a work account and a recording everyone knew about. The quotes make every finding checkable against the transcript, and permission to report nothing makes the assistant less likely to invent a fault to fill the list. Check each quote before you act on it, because automatic transcripts sometimes put a line against the wrong speaker.

Meeting transcript feedback prompt. A meeting transcript feedback prompt asks an AI assistant to read the transcript of a meeting you took part in and report, with quoted lines, what you did in it, rather than summarising what was decided.

Written by Tayyeb Mahmud, founder and Principal Architect. Claude Certified Architect – Professional · LinkedIn profile of Tayyeb Mahmud

1AYM is an OpenAI Select Partner.

Checked 11 October 2026. Anthropic’s and OpenAI’s own pages on chat, memory, projects, skills, prompting and workspace data, listed under Sources, were read on this date. Plans and menus change, so check them in your own account. This page is not legal advice.

The transcript as a mirror · 48 seconds, with captions

Why the minutes miss the part about you

Minutes are written for the people who need the outcome: what was decided, who owns each action and when the group meets again. An AI summary does that job faster, and it is still the same job. Neither tells the person chairing how they ran the meeting: the question from finance that got a vague answer, the moment the discussion moved on before anyone agreed, the planner who never got a word in.

That feedback rarely arrives on its own. People do not usually tell the chair which question was dodged, and the chair is the person least placed to notice, because they were busy running the meeting. A transcript has none of those limits. It holds every line in order with a speaker against it, so the evidence of how you chaired is already sitting in the file.

So use it as a mirror. Point the assistant at your own lines, ask for specific behaviours instead of a verdict, and make it quote its evidence. Minutes show the meeting. The transcript shows you.

Three conditions before you paste a transcript

A transcript holds other people’s words as well as yours, so the prompt only belongs in a setting that treats them with care. These three conditions come first, and none of them is optional.

A work account: Use the Claude or ChatGPT account your employer provides, under its settings. Anthropic lists no model training by default on its Team and Enterprise plans, and OpenAI says it does not use content from ChatGPT Business, Enterprise or Edu workspaces to improve its models by default [6, 9]. A personal or consumer account is the wrong place for a work meeting, and nothing confidential goes into one.
A recording everyone knew about: Only use the transcript of a meeting that everyone knew was being recorded or transcribed. Where your organisation has rules on recordings, transcripts or AI tools, those rules come before this page.
Pointed at yourself: Every request in the prompt is about the speaker you name. “Who barely spoke” is there so you can bring that person in next time, and it says nothing about their work. Using transcripts to judge colleagues is a different exercise, with HR and data-protection questions attached, and this prompt is not built for it.

The prompt to copy, word for word

Open a new chat, paste the transcript or attach the file, and say which speaker is you. Use the label exactly as the transcript writes it, for example “I am Speaker 2 in this transcript.” Then send the prompt below, unchanged.

The transcript mirror prompt

Do not summarise. List each question I ducked, where I changed the subject, and who barely spoke. Quote the lines. If none, say so.

A useful answer is short. Each question you ducked comes back with the line that asked it and the line you replied with, each change of subject with the line that changed it, and anyone who barely spoke by name or label. If the meeting was clean, it says so in a sentence.

Read it with the transcript open beside it. Search for each quoted line, check that the speaker is right, and only then decide whether the finding is fair.

Why each line of the prompt is there

The prompt is four short sentences, and each one closes off a way the answer can go wrong.

“Do not summarise”: Give an assistant a transcript without instructions and a summary is the likely result, which is the minutes again. Ruling it out sends the effort to the part you cannot get anywhere else.
Three named behaviours: A ducked question, a change of subject and a quiet voice can each be found on the page and pointed to. “How did I do?” invites a verdict and polite generalities. OpenAI’s prompting guidance points the same way: start a prompt with the result you need [3].
“Quote the lines”: A quoted line is evidence you can search for. Anthropic’s guidance on reducing hallucinations recommends pulling word-for-word quotes from long documents and finding a supporting quote for every claim, dropping any claim without one [1]. OpenAI tells users to verify names and quotes in a first result before relying on it [2].
“If none, say so”: Ask for faults and you invite faults. Anthropic’s guidance on review agents notes that a reviewer asked to find problems will usually find some [5]. Both vendors recommend the counterweight: Anthropic says to give Claude explicit permission to say it does not know, and OpenAI says an explicit gap is easier to deal with than a confident guess [1, 2]. Permission to report nothing is what makes a clean result believable.
The word “I”: Every request is about the speaker you named, which keeps the answer on the one person in the meeting you can change.

Two variations: several meetings, and the replies you owe

The prompt works on one meeting. Two changes make it more useful: running it over several meetings to find a habit, and turning what it finds into something you do next.

Variation: several meetings

I am [your speaker label] in each of these transcripts. Do not summarise. Across all of them, list each question I ducked, where I changed the subject, and who barely spoke. Group the findings by pattern and mark any pattern that appears in more than one meeting. Quote the lines and name the meeting each one comes from. If none, say so.

One meeting catches a slip. Several show a habit, and the habit is what is worth changing. Keep the transcripts of the meetings you chaired over a few weeks in one project and run the variation there; the next section says where projects sit in each product. Check the quotes the same way, because a pattern built on one misattributed line is not a pattern.

The second variation is a follow-up in the same chat, after the quotes are checked: “Turn each question I ducked into a short note to the person who asked it: their question, quoted, and a blank line for my answer. Do not write the answers.” It turns the mirror into a list of replies you owe, and it leaves the answers to you, because the assistant does not know them.

Where to run it in Claude and ChatGPT

Both products run the prompt in an ordinary chat, with no special feature switched on: paste the transcript or attach the file, then send the prompt. What differs is the settings around the chat, and for a transcript those matter more than the prompt.

Where the transcript prompt runs in Claude and ChatGPT, and the settings that matter for a meeting transcript
Dimension	Claude	ChatGPT
Where to run it	An ordinary conversation, on web, desktop or mobile. Chat is on the Free plan, and Pro and Max accounts moved to the merged Claude have one conversation with no mode to choose	Chat, which OpenAI describes as the mode for a question or back-and-forth, and which it says remains available in every workspace configuration
Training on what you paste	None by default on Team and Enterprise. Free, Pro and Max list model training as opt-out	OpenAI says it does not use content from Business, Enterprise or Edu workspaces to improve its models by default
Memory	On by default on Free, Pro and Max. On Team and Enterprise the owner decides whether it is available, and it is off for each member until they turn it on. It can be switched off for a single chat from the + menu before the first message	With Memory on, ChatGPT can remember relevant details from your chats. It is managed under Settings, Personalization, Memory
Several meetings together	A project, whose instructions apply only to the chats inside it. Projects are on all plans, and Free accounts can create up to five	A project, which can hold chats started in Chat or ChatGPT Work, with instructions that apply across its chats
The same prompt every week	Save it as a skill: turn on the skill-creator skill under Customize > Skills and ask Claude to use it to draft one	On eligible workspace accounts, save it as a skill from Plugins, then the Skills tab, then Create with chat, or ask ChatGPT in a chat to create one

The Claude rows are from Anthropic’s pricing page, its help centre articles on the merged Claude, memory and projects, and its guide to building skills [6, 7, 10, 12, 14]. The ChatGPT rows are from OpenAI’s guide to using ChatGPT, its Help Center articles on Chat and Work, memory and skills, its announcement of dots and its projects documentation [2, 8, 9, 11, 13, 15]. All were checked on 11 October 2026.

Two settings deserve a deliberate choice. In Claude, where memory is on, switch it off for this chat before the first message, since Claude saves memory as topics while you chat [10] and a transcript is mostly other people’s words. In ChatGPT, check what Memory is set to under Settings, Personalization, Memory before you paste [11].

Save the prompt as a skill only once its results are ones you trust. OpenAI’s advice for recurring work is to refine a prompt in a normal chat first [3], and the same order works in Claude.

What goes wrong, and what a transcript cannot show

The prompt is the reliable part. The failures sit mostly in the transcript and in what you do with the answer.

Speakers get muddled: Automatic transcription sometimes puts a line against the wrong person, and the assistant can only work from the labels it is given. A ducked question pinned on you that someone else answered is a false finding. Check each quote, and the name against it, in the transcript itself.
A quote can still be wrong: Asking for quotes lowers the risk of an invented finding without removing it: Anthropic says these techniques reduce hallucinations but do not eliminate them [1]. Its Claude Cowork course treats a date, name or quote you cannot trace to an input as a flag [4], and OpenAI’s guidance is that a polished response can still be incomplete or wrong [2].
The words without the room: A transcript has no tone, no faces and nothing said before the call started. A question can look ducked on the page because it was answered in a document the day before. Treat each finding as a lead to check, and decide for yourself what it means.
One meeting is a small sample: A single transcript shows a slip. Call it a habit only when the same finding turns up across several meetings.
The wrong account: The avoidable mistake is pasting a work meeting into a personal account because that tab was open. Anthropic’s training defaults differ between its personal and work plans, as the table shows, and OpenAI’s statement covers its work plans only, so a work meeting stays in the work account. Where the company has not yet chosen a work account, our comparison of the Claude and ChatGPT work plans sets out each plan’s data terms.

Where 1AYM fits

The prompt is the easy part. What decides whether your managers can use it safely is the workspace around it: which account they are in, what it keeps, whether memory is on, and who may turn a prompt that works into a skill colleagues can share. We roll out ChatGPT Enterprise and Claude Enterprise, and we build those settings as controls under our AI governance implementation work, so the rules on transcripts are set in the workspace rather than left to each person’s judgement. A fixed-scope rollout can start within a day of the scope being signed. If that would help, book a call from the end of this page or email enquiries@1aym.com.

For engineers: checking the quotes and speaker labels, and testing it before a team relies on it

The prompt is plain text, so the engineering is in checking what comes back and in running it safely across many people. Vendor guidance below is from the pages listed under Sources, checked 11 October 2026.

Ask for rows: Ask for one row per finding: the speaker label, the timestamp where the transcript has one, the quoted line and the type of finding (ducked question, change of subject, barely spoke). Rows are quicker to check than prose and can be compared across meetings.
Match quotes mechanically: Check each quoted line against the transcript with an exact string search after normalising whitespace and straight and curly quotation marks. Drop any quote that does not match rather than repairing it, then confirm the speaker label on the matching line, because a real quote on the wrong speaker is still a wrong finding.
Two passes on long transcripts: Anthropic suggests that for documents over about 20,000 tokens Claude should pull word-for-word quotes first and work only from them [1]. Here that means one pass that lists every question put to your speaker label, quoted with its position, and a second that marks each one answered, deferred or ducked, quoting the reply.
Define what counts: Anthropic’s guidance on review agents notes that a reviewer asked to find problems will usually find some [5], so keep “If none, say so” in every version and write down what counts: for example, a question addressed to you with no answer, a deferral with no owner, or a reply on a different topic.
A small test set: Before a team relies on the prompt or a skill built from it, keep three or four transcripts where you already know the answer, including one where the honest answer is none, and rerun them after every change. OpenAI gives the same advice for plugins: test on a document you know well, then on one with missing details, and rerun saved test prompts after each change [17].
Shared across a company: OpenAI’s admin checklist before a skill goes company-wide asks for a confirmed source, an accountable owner, the intended audience and a review date, a test with non-sensitive data and the least access it needs, and a named owner for re-review and retirement [16]. For a skill that reads meeting transcripts, the non-sensitive test data is the part to plan for.
Retention: A pasted transcript becomes part of the chat, so whatever retention your workspace applies to chats now applies to that copy of the meeting. Agree that with whoever owns your rules on recordings before the prompt becomes routine.

Sources

Questions about using meeting transcripts with AI

Can ChatGPT or Claude analyse a meeting transcript?

Yes. Paste the transcript into an ordinary chat or attach the file; no special feature is needed. Say which speaker is you and ask for specific findings with quoted lines rather than a summary, so you can check each one against the transcript. Use a work account and only a recording everyone knew about.

What is a good AI prompt for feedback on a meeting I ran?

Say which speaker is you, then send: “Do not summarise. List each question I ducked, where I changed the subject, and who barely spoke. Quote the lines. If none, say so.” It asks for behaviours you can verify, makes the assistant show its evidence for each one and allows a clean result.

Is it safe to put a meeting transcript into ChatGPT or Claude?

Use your employer’s account and its settings. OpenAI says it does not use content from Business, Enterprise or Edu workspaces to improve its models by default, and Anthropic lists no model training by default on Team and Enterprise. Use only a recording everyone knew about, and follow your organisation’s rules on recordings. This is not legal advice.

Why does the AI attribute a quote to the wrong speaker?

Either the transcript put the line against the wrong person, which automatic transcription sometimes does, or the assistant misread the labels. The fix is the same: search the transcript for each quoted line and check the speaker before you act on the finding.

Can I use AI on meeting transcripts to assess my team?

This prompt is built for reviewing yourself. A transcript is other people’s words, and using it to judge colleagues raises HR and data-protection questions a chat prompt does not answer. “Who barely spoke” is in the prompt so the person chairing can bring that colleague in next time.

How many meetings does it take to see a pattern?

There is no fixed number. One meeting shows a slip, and the same finding across several meetings is a habit worth changing. Put the transcripts in one project and ask which findings recur, with a quoted line from each meeting.

Can I save the prompt so I do not have to retype it?

Yes, as a skill. In Claude, turn on the skill-creator skill and ask Claude to use it to draft one. In ChatGPT, eligible workspace accounts create one from Plugins, the Skills tab and Create with chat. Refine the prompt in a normal chat until you trust its results before you save it.