If you paste discovery call notes into ChatGPT and ask it to write a proposal, you already know the draft you get back. It skims the notes. It fills the gaps by guessing. It opens with a line nobody would actually send. "I'm excited to work with you." A timeline you never agreed to. A budget the client never named.
The notes are rarely the problem. The ask is. Call notes have half-finished thoughts, casual asides, and holes you meant to follow up on. A one-shot prompt treats that mess as if it were a brief. The model then has to invent the missing pieces, and it will not tell you which ones it invented.
Weak prompt vs strong prompt
This is the weak version, and it is the one most people try first:
Write a proposal based on these discovery notes: [PASTE NOTES]
That usually returns a template with the client's name swapped in. Anything you forgot to write down gets filled in as if you had agreed it.
The stronger version is two steps in the same chat. First, make the model sort the notes and list the holes. Read that extract. Correct anything that is wrong. Only then ask it to draft.
Paste Step 1 first. Do not skip the review. Then paste Step 2 in the same thread, so the extract is already in context.
Step 1: Extract the essentials
Run this right after the call, while the conversation is still fresh. Paste your notes and, if you have it, a chunk of the client's website copy. Fill the [PASTE] bracket before you send.
Here are my discovery call notes and the client's website text: [PASTE]. Extract: goals, problems, constraints, budget signals, decision-makers and deadline. Then list what's missing that I should confirm.
What you want back is a clean list, not a draft: goals, problems, constraints, budget signals, decision-makers, deadline, then a short list of things you still need to confirm.
If the extract invents a deadline or a budget, fix it in the chat before you move on. "We never discussed a number, only that they want to keep this as a small project." That correction is the point of splitting the work. A bad assumption is much easier to catch in a list than inside a polished proposal.
Step 2: Draft the proposal
Stay in the same chat. Fill the brackets first. Use your own section list if you already have one. If you do not, the fallback in the prompt is fine. Set a word count so the first pass stays readable. 500 or 600 words is enough.
Draft the full proposal using this structure: [YOUR PROPOSAL SECTIONS, or 'situation, goals, approach, deliverables, timeline, investment, next steps']. Use the client's wording where possible. Max [WORD COUNT] words.
If you already have packages or a rate, paste those in the same message as Step 2 so the draft does not invent numbers. If the call never got to money, leave the investment section for yourself, or tell the model to keep it as a placeholder. A blank is better than a made-up figure.
A worked example (this one is made up)
Here is a fictional set of call notes, to show what the two steps produce. There is no real client named Harbor Street Bakery.
Sample call notes (fictional):
Call with Maya Chen, owner of Harbor Street Bakery, a made-up neighborhood bakery in Austin. They still take catering orders by Instagram DM and a shared inbox. Maya wants a simple online ordering page for catering and weekly bun subscriptions, not a full e-commerce rebuild. Holiday tray season starts the first week of December, and she wants something live before then if possible. Her brother Jordan handles the books and must approve any spend. She said they "don't have a huge budget" and asked whether this could be a small project rather than a retainer. Current site is a one-page Squarespace with hours, photos, and a contact form to Maya's personal email. Pickup only, no delivery. She wants it "to feel like the shop, warm, not corporate." I forgot to ask about weekly order volume and who writes the product copy.
What Step 1 might extract:
- Goals: a simple ordering page for catering and weekly bun subscriptions, live before early December if possible. Should feel warm and like the shop, not corporate.
- Problems: orders arrive through Instagram DMs and a shared inbox. The Squarespace page is brochure-only. The contact form lands in Maya's personal email.
- Constraints: pickup only. Not a full e-commerce rebuild. She wants a small project, not a retainer.
- Budget signals: "don't have a huge budget." No number given. Jordan (brother, handles the books) must approve spend.
- Decision-makers: Maya Chen (owner) and Jordan (spend approval).
- Deadline: she wants it live before holiday tray season in early December. No hard date was locked in.
- Missing, confirm before drafting: weekly order volume, who writes product copy, a budget range, whether early December is a hard deadline, what the ordering page must include, and whether Jordan will join a follow-up.
The extract does not turn "don't have a huge budget" into a number. That is what you want. If anything in the list is off, correct it in plain language before Step 2.
A short excerpt of what Step 2 might draft, with the word count set to 500 and the default sections left in place:
Situation
Harbor Street Bakery still takes catering and subscription orders through Instagram DMs and a shared inbox. The current Squarespace site covers hours, photos, and a contact form, but it does not take orders. Maya wants a simple ordering page that feels like the shop, warm rather than corporate, ready before holiday tray season if we can manage it.Goals
Stand up an ordering page for catering trays and weekly bun subscriptions, pickup only, without turning this into a full e-commerce rebuild.Approach
...
The rest would follow with approach, deliverables, timeline, investment, and next steps. In a real pass, paste your rate or packages into Step 2, or write the investment section yourself. The fictional notes never named a number, so leave that section as a placeholder.
The useful part is the wording. "Warm, not corporate" and "small project rather than a retainer" show up because Step 1 already pulled them out of the notes. A one-shot prompt often drops those phrases and swaps in generic agency language.
Why it works
Step 1 makes the model sort the notes and flag holes before it writes anything. It is less likely to invent a budget or a deadline you never discussed, because you asked it to list what is missing instead of filling the gaps.
Step 2 already has that extract in the chat. The draft can reuse the client's wording and follow a fixed structure, instead of opening with a generic pitch. If a section comes out thin, that usually means something was missing from the notes, not that you need a longer prompt.
These two prompts are the extract-and-draft pair from a longer proposal workflow. For a first draft the same day as the call, this is the part that changes the output the most.
Tips
Fill every bracket before you send. A prompt that still says
[PASTE]or[WORD COUNT]will stall or guess.Paste a little website copy with the notes in Step 1, so the model can match how the client already describes themselves.
Correct the extract in the chat, out loud. "Jordan approves spend, but Maya is the only decision-maker on scope." That one line will usually show up in the draft.
Do not let the model invent dates, fees, or team size. If Step 1 listed them as missing, keep them missing in Step 2, or go back to the client.
Edit the investment section yourself. Even when you paste your rates in, read that part as if a client will screenshot it.
If the first draft is stiff, ask for a rewrite of one section only, using the client's phrases from Step 1.
These two prompts come from the AI Prompt Playbook we put together at Kitwise Studio.
Top comments (0)