ChatGPT can only write a follow-up as warm as the notes you hand it.

The Thursday after the conference

It's Thursday. The conference ended Tuesday, and you've got a stack of badge scans and a notes app with three half-sentences in it. You open ChatGPT and type: "Met Dana, VP of Sales, at SaaStr. Write a warm follow-up."

The draft comes back in two seconds. "It was great connecting at SaaStr! I really enjoyed our conversation and would love to explore how we might work together." Clean grammar. Friendly tone. Interchangeable with every other follow-up landing in Dana's inbox this week.

So you ask for warmer.

You get exclamation points.

The context ceiling

A drafted follow-up can't be warmer than the context behind it. Call it the context ceiling. ChatGPT is faithful to its input: give it a name, a title and an event, and it writes the email those three facts support. Which is a polite, empty one.

The model did its job. It wrote the email the prompt allowed. Warmth in a follow-up comes from evidence that you were listening: a thing they said, a problem they mentioned, a next step you both agreed to.

None of that is on a badge. None of it sits in the address-book fields a card scanner fills in. And if you didn't write it down, the model can't supply it, and you really don't want it to try.

Why better prompts don't fix a cold follow-up

The standard advice treats this as a prompting problem. Add a persona ("you're a seasoned enterprise seller"). Specify a tone. Ask for three variations. Paste in a template from a post that promises replies.

Those moves change the register of the email. They don't add a single fact about Dana. A persona with no content produces a confident-sounding stranger, and a tone instruction with no content produces adjectives. Ask for "more personal" without giving it anything personal and the model has two options: stay generic, or make something up.

The made-up version is the dangerous one. A follow-up that references a conversation you never had is worse than a bland one, because Dana will notice.

What do you need to give ChatGPT to write a good follow-up?

The companies building these tools say much the same thing in their own documentation. Both OpenAI and Anthropic publish prompting guidance, and both point users toward specificity: give the model the details and the context of the task, and say plainly what you want. Neither offers a trick that substitutes for the details themselves.

For a conference follow-up, the details that matter fall into three buckets. Something they said, ideally close to their words. The topic that surfaced, the thread you'd pick back up. And a concrete next step: the thing you offered, or the thing you both agreed to. Add who they are and what you want from the email, and the draft has something to stand on.

Now look at where each of those lives a few days after the event.

Same model, two prompts

Take the same contact twice.

The sparse version is what most people have by Thursday: "Dana, VP Sales, met at SaaStr." The output reads like the email at the top of this piece. It could go to anyone on the expo floor.

The rich version is what you'd have if you'd spent a minute or two on it before the next session. Dana's team is hiring a batch of new AEs next quarter and ramp time is the worry. Dana called the current onboarding "a shared drive and a prayer." You offered to send the onboarding checklist your team uses. You both said a call the week after next made sense.

Same model, same instruction. The second draft opens on the ramp problem, echoes the shared-drive line with a light touch, attaches the checklist and proposes two call times. It reads like it came from the person Dana talked to.

Which it did. The model just had the receipts.

The bottleneck moved to the room

Drafting is close to solved. Capture isn't.

Anyone can produce a fluent follow-up in seconds now, so the scarce input is the three details, and they have a short shelf life. By the time you're home, the conversations from the first morning have blurred into each other. We covered that decay in [The 72-Hour Memory Cliff](/articles/72-hour-memory-cliff/).

This shifts what a capture tool should be judged on. Most card scanners compete on reading a card accurately: name, title, phone, email. That's the one row in the table above that adds no warmth, and it's the argument behind [OCR is a feature, not the product](/articles/ocr-is-a-feature-not-the-product/). The fields that make a drafted follow-up land are the ones a scanner can't read, because they were never printed on anything.

So the practical move is boring. Write the three details down while you're still standing in the hallway.

The model can't remember a conversation you didn't write down.

Where Met fits

Met is built for that hallway moment. Event Mode keeps the room captured while you keep talking, so the detail Dana dropped about onboarding gets saved with Dana's contact while the conversation is still going. When you sit down on Thursday, the context is already sitting next to the name.

From there, draft with whatever you like. Paste the notes into ChatGPT, or write the email yourself in a few minutes because you finally remember what you talked about. Met's job is the input. Your contacts stay in your iCloud, there are no ads, and nobody is buying your network from us.

The drafting tools will keep getting better at writing. They still won't have been in the room.