Why the default summary is useless
Paste a transcript, ask for a summary, and you get a paragraph that says the team discussed the project and agreed to follow up. It is accurate and it is worthless, because it strips out the only two things anyone opens meeting notes for: what was decided and who owes what.
The fix is to name the four things a meeting actually produces and ask for those. Decisions. Commitments. Open questions. Disagreements. That last one gets lost the most often and matters the most later.
Two models, two passes. Claude reads the room and writes the narrative, GPT extracts the structure without editorialising. Splitting the job takes one model switch and produces notes people actually read.
Two prompts do the whole job
Prompt 1: the recap, in Claude
Here is a meeting transcript. Produce: (1) a one-line headline stating what this meeting actually resolved; (2) the decisions made, with who made each one; (3) five to eight bullets covering the substance, in the order it matters rather than the order it was said; (4) open questions that were raised and not answered; (5) any point where people disagreed, stating both positions without resolving them. Do not include pleasantries, scheduling talk, or anything discussed and abandoned. If a decision was implied but never stated explicitly, list it separately under UNCONFIRMED. Transcript: [paste].
Prompt 2: the action items, in GPT
From the same transcript, extract every commitment as a table: task, owner, due date, and the exact quote it came from. Rules: only include items where someone actually committed, not things that were merely suggested. If the owner is ambiguous, write UNASSIGNED rather than guessing. If no date was stated, write NO DATE rather than inferring one. Then list separately any task that was raised and never picked up by anyone.
The quote column is what makes this trustworthy. It turns a disputed action item from an argument into a lookup, and it makes fabricated commitments immediately obvious because there is nothing to quote.
Prompt 3, optional: the follow-up email
Draft a follow-up email from the recap and action items above. Audience: [who was and was not in the room]. Lead with the decision. Keep it under 200 words. State each person's action item explicitly so nobody has to interpret. Tone: [your voice]. Flag anything I should confirm before sending.
Templates for the meetings you actually have
Different meetings produce different artifacts. Adapt the extraction, not just the wording.
| Meeting type | Ask for | The thing people forget |
|---|---|---|
| Customer or sales call | Objections raised, verbatim language for the problem, what they compared you to, next step and who owns it | Copy the customer's exact words rather than paraphrasing them |
| Decision meeting | The decision, the alternatives rejected and why, who has to be told | Recording the rejected options, which is what prevents relitigating it in six weeks |
| Standup or status | Blockers with owners, changes to previous estimates, anything that slipped silently | Slippage that nobody said out loud |
| User interview | Jobs they described, workarounds they built, exact frustration quotes, contradictions between what they said and did | Contradictions, which are the most useful part |
| Interview or hiring panel | Evidence against each competency, with quotes, and questions left unasked | Separating evidence from impression |
| Board or investor meeting | Commitments made by you, questions you could not answer, follow-ups requested | The questions you could not answer |
Keep whichever of these you run weekly as a prompt you paste, with a {{transcript}} placeholder marking where the transcript goes, so the whole workflow becomes paste, run, skim.
How do you get a usable transcript?
Whizi does not join or record your calls. You bring the transcript, and every major platform will produce one: Zoom, Teams, and Meet all have built-in transcription, and dedicated notetakers export plain text.
A few things that materially improve the output:
- Keep speaker labels. Attribution is most of the value. A transcript with speakers stripped can still be summarized, but every action item becomes UNASSIGNED.
- Do not clean it up first. Filler words and false starts cost you nothing and the model ignores them. Time spent tidying a transcript is wasted.
- Upload long transcripts rather than pasting them. A two hour meeting is a lot of text. Uploading it as a file keeps the conversation readable, and Gemini 3.5 Flash reads up to 1M tokens, about 1,900 manuscript pages, in one pass.
- Include the agenda if you have one. It tells the model what was supposed to happen, which makes the "raised and never resolved" section far more accurate.
What to check before you send it
Three specific failure modes, all of which have a cheap check.
Invented commitments. A model will occasionally convert "someone should probably look at that" into an owned action item with a name attached. The quote column catches this instantly, which is why it is in the prompt.
Flattened disagreement. Summaries drift toward consensus, because most text in the world resolves. If two people disagreed and the notes read as though everyone agreed, the notes are wrong in the way that causes trouble two weeks later. Ask explicitly: Where did people disagree, and was it resolved?
Wrong attribution on a crosstalk-heavy call. Transcription tools mislabel speakers when people talk over each other. Spot check the attribution on anything consequential before it goes into a document with someone's name on it.
A useful last pass: What is in this transcript that a busy person would need to know but would not find in the summary above? It regularly surfaces one thing worth adding.
- Keep speaker labels in the transcript, since attribution is most of the value
- Ask for decisions, commitments, open questions, and disagreements rather than a summary
- Require the exact supporting quote on every extracted action item
- Use UNASSIGNED and NO DATE instead of letting the model guess
- Check the disagreement section, since summaries drift toward false consensus
- Keep a reusable prompt for your most frequent meeting type
Frequently asked questions
Does Whizi record meetings?
No. Whizi does not join calls or capture audio. You bring the transcript from whatever you already use, whether that is Zoom, Teams, Google Meet, or a dedicated notetaker, and paste or upload it. Keeping speaker labels intact matters more than the source, because attribution is where most of the value in a meeting summary comes from.
Is this safe for sensitive meetings?
Whizi does not train on your conversations and each provider's data policy can be reviewed before you enable that model, but a transcript is one of the most sensitive documents a company produces. For legal, HR, board, and personnel matters, check your internal policy first, and consider summarizing from your own notes rather than the raw transcript. Stripping names and identifiers you do not need before uploading is a cheap extra safeguard.
Which model should I use?
Claude for the narrative recap, because it reads intent well and writes notes people finish reading. GPT for the action item extraction, because it holds a strict output format without editorialising. Gemini when the transcript is very long, since it can take a multi-hour meeting in one pass without chunking. On credits, the split is cheap to run: GPT-5.6 Terra costs 4 credits a message, Gemini 3.5 Flash 8, and Claude Sonnet 5 10, per Whizi's model cost index, so both passes on one transcript come to 14 credits.
How long can the transcript be?
Longer than most meetings produce. A typical hour-long call is well within range for any of the models. For multi-hour sessions, all-day workshops, or a batch of several meetings at once, upload the files and use Gemini 3.5 Flash, whose 1M token context, roughly 1,900 manuscript pages, holds the whole set at once.
Can it tell me what was decided if nobody said it clearly?
It can flag it, and that is the right behaviour. The prompt on this page asks the model to list implied but unstated decisions separately under UNCONFIRMED, which surfaces them for you to confirm rather than recording a decision that was never actually made. Meetings where nothing was explicitly decided are extremely common, and notes that pretend otherwise cause the problem later.