Daytapes

Guides · AI

How to ask AI good questions about a recording someone shared

Ask an AI about a recording someone shared with you: when to ask the sender first, prompts that keep answers tied to the transcript, and what AI gets wrong.

By Daytapes · Updated · Research-based

Ask specific questions, ask for word-for-word quotes with timestamps, and let the AI say "that's not in the recording." Then play the moment it points to before you rely on it. And because the recording isn't yours, use the page's own Ask box if it has one, or ask the sender before you paste their recording into another AI service.

This guide is for the person on the receiving end: someone sent you a meeting, a lesson, a family story or a call, and you want help finding what matters in it. The advice on prompts and mistakes comes from the AI makers' own documentation and published research (sources below), not from our own testing.

First: whose recording is it?

A shared recording belongs to the person who sent it and to everyone talking in it. They agreed to share it with you, not with every service you use. Pasting a transcript into ChatGPT, Claude, Gemini or Grok sends it to that company, where it's stored and handled under your account's data settings.

So, in order:

  1. Use the Ask box on the page if there is one. On Daytapes, a shared library has Ask this library, and a vault has Ask about these recordings. The words stay on Daytapes, and the answer comes only from the recordings you were sent.
  2. If there's no Ask box, or you want your own assistant, ask the sender: "Mind if I run this through Claude to pull out the dates?" Most people say yes. Some recordings, like a doctor's visit or a family legal matter, deserve a no.
  3. Paste only the part you need. If you want the action items from the last ten minutes, paste the last ten minutes.
  4. Check your assistant's data settings first. For example, Claude has a "Help improve our AI models" switch under Settings › Privacy that controls whether new chats can be used for model training (per Anthropic's Privacy Center, as of November 18, 2026). Our guide to using AI with your recordings covers ChatGPT and Grok too.

How the Ask box on a Daytapes page works

The Ask box on a library or vault first finds the moments that match the words in your question, then answers only from those moments, with a number after each fact that links to the exact spot. If nothing matches, it says "I couldn't find that in these recordings" instead of guessing.

Two things follow from that:

  • Use the words people actually said. "What did Maya say about the deposit?" works better than "What happened with money?" because "deposit" is likely in the transcript.
  • Parts that weren't shared with you don't exist for it. If the sender kept a stretch out of your copy, that text was never on your page. Ask can't find it, and neither can any AI you paste the page into.

A single shared recording doesn't have an Ask box. You can still read the transcript as it plays and use Copy link to this moment to point at a spot.

Prompts that work

Anthropic's documentation suggests three habits that apply to any assistant: give the AI permission to say it doesn't know, which it says "can drastically reduce false information"; for long documents, ask for "word-for-word quotes first" before the answer; and put long material first with your question at the end, which it says can improve answers "by up to 30 percent in tests."

Put together, a good starting prompt looks like this:

Here is a transcript of a recording someone shared with me. [paste the transcript]

Question: What dates or deadlines were agreed on? Answer with a word-for-word quote and the timestamp for each one. If the transcript doesn't say, write "Not in the transcript." Don't guess.

Some questions that tend to get useful, checkable answers:

  • "What was decided, and by whom? Quote each decision with its timestamp."
  • "List every number mentioned (prices, dates, measurements) with the sentence it's in."
  • "What did Jordan ask for? Quote it." Use the name the transcript shows.
  • "What questions were raised but not answered?"
  • "Summarize the part from 12:00 to 20:00 in five bullet points, each with a quote."
  • "Explain this term the way the speaker used it." Good for lectures and classes.

Questions that tend to go wrong:

  • "What did they really mean?" The AI can only see words, not tone or history. You'll get a confident guess.
  • "Who was right?" That's your judgment to make, not a summary.
  • "Summarize this two-hour meeting" in one shot. Ask about one topic or one stretch at a time instead (see the next section).

What AI gets wrong

It states things that aren't there

NIST's Generative AI Profile defines confabulation as "the production of confidently stated but erroneous or false content (known colloquially as 'hallucinations' or 'fabrications')." OpenAI's researchers wrote in September 2025 that models do this partly because "training and evaluation procedures reward guessing over acknowledging uncertainty." Google's Gemini help page says plainly: "Gemini Apps can make mistakes. When using Gemini Apps, double-check responses."

It happens with conversations too. TofuEval, a 2024 study of AI summaries of dialogues, found that "existing LLMs hallucinate significant amounts of factual errors in the dialogue domain, regardless of the model's size."

What to do: ask for quotes and timestamps, and play the moment. If the quote isn't there, the answer isn't either.

It misses the middle

A 2024 study, "Lost in the Middle," found that language models do best "when relevant information occurs at the beginning or end of the input context," and do worse when it's in the middle of a long one. A long meeting has a lot of middle.

What to do: ask about one stretch at a time, or about one topic using words from the recording. "What did they say about the roof?" beats "What were the main points?"

The transcript itself can be wrong

The AI is reading a transcript, and the transcript is a machine's best hearing of the audio. Names, numbers and jargon are the usual misses. Who said what can be wrong too: AssemblyAI's documentation on speaker labels says its model "may struggle to create separate clusters for speakers who only contribute short phrases like 'Yeah', 'Right', or 'Sounds good'," and that "overlapping speech between speakers can reduce diarization accuracy."

What to do: before you quote someone, or repeat a number, play that moment. If a name looks wrong in the transcript, the AI's answer will carry the same mistake.

It fills gaps with what's typical

If the recording doesn't say when the next meeting is, an AI may suggest "next week" because that's common. Telling it to write "Not in the transcript" when the answer isn't there makes this less likely, but not impossible.

A worked example (made-up names)

Alex's sister Maya records their parents' planning call about a fall family reunion and shares a vault with Alex, who missed the call. Alex wants the dates and who's bringing what.

  1. Alex opens the vault and types into Ask about these recordings: "Which weekend was picked for the reunion?" The answer gives a date with a [1] after it. Alex taps [1], and the player jumps to the moment Dad says it.
  2. Next: "Who is bringing the tables?" The answer says "I couldn't find that in these recordings." Alex tries "tables" in search, finds nothing, and texts Maya instead.
  3. Alex wants a checklist, so Alex asks Maya, "Okay if I paste the last ten minutes into Claude?" Maya says yes. Alex pastes only that stretch, with the prompt above. Two of the five items come back with quotes; one says "Not in the transcript," so Alex leaves it off the list.

Before you share what the AI told you

  • Share the moment, not the summary. A Copy link to this moment link lets people hear it themselves, and only works for people who already have access.
  • Don't forward AI summaries of someone else's recording without asking. The summary carries the same private details.
  • Say it's an AI summary when you pass it on, so people know to check it.

Questions

Can I paste a recording someone shared with me into ChatGPT or Claude?

Ask the sender first. The recording is theirs and the other people's in it, and pasting it into an AI service sends it to another company under that company's data settings. If the share page has an Ask box, use it: it stays on the page.

Why did the AI say something that isn't in the recording?

AI models can state false things confidently. NIST calls this confabulation. Ask for word-for-word quotes with timestamps, and let the AI say it doesn't know.

Can the AI see the parts the sender didn't share with me?

Not on Daytapes. The Ask box on a library or vault only searches the words you're allowed to hear, and those parts were never on your page, so there's nothing for it or any other AI to read.

Is the speaker shown in a transcript always right?

No. Speech-to-text services say telling voices apart gets harder with short replies, cross-talk and background noise. Check a quote by playing that moment before you repeat who said it.

Sources

  1. Anthropic docs: Reduce hallucinations (accessed November 18, 2026)
  2. Anthropic docs: Prompting best practices (long context tips) (accessed November 18, 2026)
  3. Kalai et al., Why Language Models Hallucinate (OpenAI, September 2025) (accessed November 18, 2026)
  4. Google Gemini Apps Help: Use Gemini Apps (accessed November 18, 2026)
  5. NIST AI 600-1: Generative AI Profile (July 2024) (accessed November 18, 2026)
  6. Liu et al., Lost in the Middle: How Language Models Use Long Contexts (TACL, 2024) (accessed November 18, 2026)
  7. Tang et al., TofuEval: Evaluating Hallucinations of LLMs on Topic-Focused Dialogue Summarization (NAACL 2024) (accessed November 18, 2026)
  8. AssemblyAI docs: Speaker diarization (accessed November 18, 2026)
  9. Anthropic Privacy Center: How do I change my model improvement privacy settings? (accessed November 18, 2026)