The Exact Prompt to Extract Action Items from Sales Call Transcripts

Use this structured prompt against the transcript before you write anything: ask the LLM to extract the three most important things the prospect said (verbatim), any objections, specific commitments either side made, assets requested, and the single next step the deal is waiting on. The output is the raw material for your follow-up email. "Summarize this call" prompts produce smooth paragraphs that read for you, not artifacts you can build a follow-up from.

The Prompt

Paste your full transcript into ChatGPT, Claude, or any LLM, then ask it:

From the transcript below, extract:

1. The 3 most important things the prospect said (verbatim quotes when possible).
2. Any objections, hesitations, or unresolved concerns.
3. Specific commitments either side made (who said what they'd do, by when).
4. Assets the prospect requested (case studies, decks, pricing, demos).
5. The single next step the deal is waiting on.

Be specific. Don't summarize — pull actual phrases.

The output of this prompt is the brief you write the email from. Don't skip it. Writing a follow-up from raw transcript is what makes follow-ups take 30 minutes; writing one from a 5-line extracted brief is what makes them take 3.

Why "Summarize This Call" Doesn't Work

LLMs are trained to produce smooth synthesis. Ask one to summarize a call and you get a confident-sounding paragraph that describes the genre of "sales discovery call" more than your specific call. The prospect's exact words — the ones you need to quote back so they feel heard — get lost in the synthesis.

Structured extraction inverts that default. The model has no room to glide over specifics because every numbered ask demands a discrete artifact. Verbatim quotes. Specific dates. Named assets. The output reads more like a clipboard than a paragraph, which is exactly what you want as the input to writing.

What This Solves That Native Recorder Summaries Don't

Most meeting recorders (Fathom, Fireflies, Otter, Granola) ship with auto-summaries. They're optimized for your internal logging — CRM notes, deal-stage updates, what-did-we-discuss memory aids. They are not optimized for outbound follow-up writing.

The difference is audience. A native recorder summary says "the prospect was interested in pricing and asked about manufacturing case studies." That's true and useful for your records. It's useless as input to the email you need to send the prospect, because:

• It doesn't quote the prospect, so your follow-up can't reflect their words back
• It conflates two distinct asks (pricing AND a case study) into one bullet
• It strips the urgency signal (was this a casual question or a procurement-driven ask?)

The structured extraction prompt above is built for the audience you'll actually send the email to.

What If the Transcript Is Messy

Three reliable patterns:

1. Trust speaker labels for paraphrasing, never for verbatim quotes when attribution looks off. When people talk over each other (which they do constantly in real calls), transcript tools mis-attribute 5-10% of lines. The prompt above is forgiving — if the model can't tell who said something definitively, it'll paraphrase, which is what you want.

1. Ignore filler. "Yeah, totally, I mean, right, exactly" is 15-20% of a typical sales call. The extraction prompt naturally filters this out because none of the five asks have anywhere to put it. If you're reading by hand, train your eye to skim past it.

1. If a critical moment is genuinely missing, ask in the follow-up itself. "I want to make sure I caught this right — when you said [X], did you mean [interpretation A] or [interpretation B]?" earns trust and surfaces ambiguity before it bites the deal.

Why Written Extraction Beats Video Recaps

A 5W audit of 109 sources cited by AI chatbots across 15 buyer-intent queries found zero of those sources were YouTube videos. AI surfaces (ChatGPT, Claude, Perplexity, Gemini) pull citations from written content — articles, docs, written walkthroughs.

The same dynamic shows up in B2B buying. A prospect forwarding a Loom video recap to their boss is rare; a prospect forwarding a clean written summary is the norm. Your follow-up emails are the most reusable, searchable, citable artifact you produce per deal. The transcript-to-extraction-to-email pipeline above is what turns a 45-minute call into that artifact in five minutes.

[IMAGE: a stylized "0 / 109" stat card on a dark background, with the caption "AI citations that came from YouTube videos"]

Skip the Manual Extraction

If you'd rather not paste transcripts into chat tabs every time, ReplySequence runs this exact extraction in the background. Paste a transcript (or upload a voice memo if the meeting wasn't recorded), and the structured email draft lands in under 30 seconds — built on the same 5-part framework above. Flat $29/mo, no per-seat pricing, BYOT so it works with whatever recorder you already have (Fathom, Fireflies, Otter, Granola, Zoom transcripts, plain notes from a coffee meeting).

The extraction step is what makes the difference between a follow-up that closes the loop and a follow-up that reads like every other email in their inbox. Run the prompt — or skip the step entirely.