We write proposals for a living and used to lose most of a morning to each one. Not because writing is hard, but because we would re-listen to the call, try to remember what the client actually said, and end up describing our own services instead of their problem. The proposals that win are the ones where the client reads the first paragraph and thinks "yes, that's it."
The transcript already contains that paragraph. You just have to get it out.
Record the call and get a transcript
You need two things: a recording of the call with a usable transcript, and a model with a large enough context window to read the whole thing.
For the recording we use Granola, because it sits on the call without joining it as a bot participant, which matters when you are talking to someone who has never been on a recorded call before. Any transcription tool works.
Get permission before recording. (In the margin: always ask.) In Canada one-party consent covers the legality, but a client who finds out afterwards is a client you have annoyed for no reason. We ask "mind if we record this so we don't have to take notes?" and nobody has ever said no.
Two more setup notes:
- Strip names and anything sensitive if the call covered patient data, payroll, or anything under an NDA. A find-and-replace before pasting takes ten seconds.
- Keep a file of your last three won proposals. You will use it in the Variations section below.
Pro tip
Do not feed the tool's own AI summary into the next step. The summary has already thrown away the exact phrasing, and the exact phrasing is the whole point.
Checkpoint
Plain text you can paste, not a summary.
Pull a brief out of the transcript
the step that pays off
This runs in two passes, and the split is doing real work. If you ask for a proposal directly from the transcript, the model compresses everything on the way through and you get a generic document that could have been written without the call. Extracting a structured brief first forces the quotes to survive.
Pass one pulls the brief.
Pull a brief out of the transcript
Checkpoint
A brief with the problem in their words, as verbatim quotes.
Draft the proposal from the brief
Pass two writes from the brief.
The instruction that changed our output most is the
[ASK: ...]rule in pass two. Before we added it, the model would quietly invent a headcount or a launch date to fill a gap, and we would not notice until the client did. Now the gaps are visible in the draft and we either fill them from memory or send a two-line email.Draft the proposal from the brief
Pro tip
Propose one scope, not three tiers. Tiers make people shop instead of decide.
Checkpoint
A proposal draft that is roughly 80% there.
Finish it by hand
What we still do by hand, every time:
- Set the price. The model does not know what we are worth or how much we want this project. It puts a number in as a placeholder and we replace it.
- Cut a phase. First drafts always propose one phase too many. The work reads more confident when it is smaller. (In the margin: one phase fewer.)
- Rewrite the opening line. It is usually accurate and flat. This is the sentence the client reads, so it is worth the two minutes.
Checkpoint
A draft with your price in it, one phase fewer, and an opening line you wrote.
The takeaway
Time went from roughly 40 minutes to roughly 10, and the drafts are better than the ones we wrote tired at 6pm. The improvement is not the writing. It is that the client's own words end up in the document instead of our paraphrase of them.
One honest caveat: this makes bad-fit projects easier to propose too. A fast proposal is only a good thing if you are still willing to say no.
Variations
Ground it in proposals that already won. Paste two or three of your past signed proposals above the brief with "match the structure and tone of these." Output quality jumps more from this than from any wording change to the prompt itself.
Follow-up email instead of a proposal. Same pass one, then ask for a 150-word recap email with three bullets and one question. Send it within an hour of the call. This alone has kept deals alive that would have gone quiet.
Discovery-question review. Run pass one on your last five calls and read only the "open questions" sections. Ours had the same gap four times out of five, so we added it to our call script.
Scope-change detection. Once a project is running, run the same extraction on a check-in call and diff it against the original brief. Anything new is either a change order or something you are about to do for free.
Don't bother with a custom GPT or a project for this yet. We tried. Two pasted prompts in a fresh conversation gave us better results than a configured assistant that slowly drifted, and there was nothing to maintain.