ChatGPT Prompts: From Meeting Notes to Campaign Plan

The messiest input in marketing is not a bad brief. It is the kickoff meeting: forty minutes of half-decisions, three people talking past each other about the audience, a budget number someone said out loud once, and a page of notes that made sense only while you were writing them. Somewhere in that page is a campaign. My job after every one of those meetings used to be excavating it by hand, and this workflow is how I handed the excavation to the model without handing over the decisions.

It is a chain of five prompts - extract, check gaps, confirm, brief, plan - and it belongs to the same family as everything in my prompt library: each prompt does one job, and I read the output between steps. The steps matter more here than almost anywhere else, because meeting notes are exactly the kind of ambiguous input a model loves to “fix” by quietly inventing what nobody said.

Why notes can’t go straight to a plan

Ask ChatGPT to turn raw notes into a campaign plan in one step and you will get a plan. It will be complete, confident and partly fiction. The notes said “maybe LinkedIn?” and the plan says LinkedIn is the primary channel. The notes never mentioned a budget and the plan has allocated one. A model given gaps fills them, and meeting notes are mostly gaps.

So my chain separates the two jobs that a one-shot prompt smears together. First, establish what was actually said - decisions, facts, open questions - with the model forbidden to add anything. Only then, once the record is confirmed by the people who were in the room, does the model get to do generative work on top of it. It is the same discipline as prompt chaining for content: the cheap human check sits early, where a wrong fact is one line to fix instead of a plan to unwind.

Prompt 1: extract what was actually said

Input is your raw notes, pasted as they are. Typos, fragments and half-sentences are fine. What the model needs is permission to leave holes.

Below are my raw notes from a campaign kickoff meeting. Sort every
point into exactly one of these buckets:

1. DECISIONS - things the group clearly agreed on
2. FACTS - concrete inputs mentioned (budget, dates, audience,
   product details, constraints)
3. OPEN QUESTIONS - things raised but not resolved
4. ACTION ITEMS - who is doing what, if named

Rules: use only what is in the notes. Do not infer, complete or
improve anything. If a point is ambiguous, put it in OPEN QUESTIONS
with a note on what is unclear. If a bucket is empty, say so.

Notes: [paste]

The bucket that earns its keep is OPEN QUESTIONS. A meeting that felt decisive routinely produces a longer list of unresolved points than of decisions, and seeing that in writing an hour after the meeting - not two weeks later, mid-campaign - is the single biggest win in this workflow. The experience is common enough to be measured: in Atlassian’s survey of 5,000 knowledge workers, 54% said they frequently leave meetings without a clear idea of next steps or who owns which task. My kickoffs were no exception until this step made the fog visible.

Prompt 2: find what the plan will need and the meeting never covered

The extraction tells you what was said. This step tells you what wasn’t, before the gaps get papered over.

Here is the sorted record of a campaign kickoff meeting: [paste
prompt 1 output]

You are about to help me turn this into a campaign brief. List the
questions you would need answered before the brief can be written,
in priority order. For each: why it matters and who in a typical
marketing team could answer it. Do not answer any of them yourself.

“Do not answer any of them yourself” is the load-bearing line. Without it the model helpfully proposes answers, and a proposed answer reads enough like a decision that it becomes one by default. I want a list of questions I can take to actual humans, because the answers are theirs to give.

Prompt 3: the confirmation message - the human gate

Before anything generative happens, the record goes back to the room. I have the model draft that message so the step actually happens instead of staying a good intention.

Turn the meeting record above into a short follow-up message to the
attendees. Structure: what we decided, what we still owe answers on
(with a named owner per question where the action items suggest
one), and one line asking people to correct anything I got wrong by
[day]. Keep it under 200 words, plain tone, no preamble.

This is the gate that makes the rest of the chain safe. Every downstream step builds on this record, so a wrong item here is a wrong campaign later. In my experience the corrections that come back are small but never zero - a date, a scope line, one “that was floated, not decided.” Each one of those is an error that would otherwise have shipped into the plan wearing a suit.

Prompt 4: the brief

Once the record is confirmed and the open questions have owners or answers, the model gets to build. The target structure is the same one I defend in the content brief workflow: a brief is a set of decisions, not a vibe.

Using only the confirmed meeting record and the answered questions
below, draft a campaign brief with these sections:

- Objective (one sentence, with the success metric)
- Audience (who, and what we know about them from the record)
- Core message (what we want them to think or do)
- Channels (only ones the record supports, with the reasoning)
- Budget and timeline (as stated; mark estimates as estimates)
- Out of scope (explicitly excluded, per the record)
- Remaining open questions (anything still unanswered)

Where the record does not support a section, write [NOT DECIDED -
NEEDS OWNER] rather than filling it in. Do not invent numbers,
channels or audience details.

Confirmed record: [paste]
Answered questions: [paste]

The [NOT DECIDED] convention carries the anti-invention rule from prompt 1 through the generative step, and I keep those markers visible in the version stakeholders see. A brief that admits its holes gets them filled. A brief that hides them gets approved, and the holes surface in week three as scope changes.

Prompt 5: the plan

The plan is the brief unpacked into sequence and ownership. This is the most mechanical step in the chain, which is exactly why it goes last: by now the inputs are solid and the model is formatting decisions rather than making them.

Turn this campaign brief into an execution plan:

1. Workstreams - group the work (creative, channels, landing pages,
   measurement, approvals)
2. Sequence - what must happen before what, as a simple ordered
   list per workstream with rough durations
3. Owners - assign from this team list: [names and roles]. Where no
   listed person fits, write [UNASSIGNED]
4. First week - the specific tasks that should start immediately
5. Risks - the three most likely ways this plan slips, based only
   on the open questions and constraints in the brief

Brief: [paste]

I ask for risks from the brief’s own open questions rather than in general, because generic risk lists are filler and the model will happily produce one. Tied to the actual unresolved points, the risk section becomes a to-do list for the first status meeting instead.

For a campaign type I run repeatedly, this whole chain lives in standing context - the team list, the house brief format and the channel constraints sit in a system prompt or a Claude Project, so each new meeting only needs the notes pasted in.

Where this workflow fails

Three failure modes, all of which I have met personally.

Garbage notes stay garbage. The chain sorts and structures what you captured. It cannot recover the decision you didn’t write down. My fix is unglamorous: during the meeting I mark decisions with a “D” as they happen, which costs nothing and makes prompt 1 dramatically more accurate. An AI notetaker transcript works as input too, with the same chain on top - though I tell the room it is running, and I still do the confirmation step, because transcription is not agreement.

The confirmation step gets skipped under deadline. It feels like process for its own sake right up until a plan built on an unconfirmed record collapses in review. When I am genuinely out of time, I shrink the gate rather than drop it - the follow-up message goes out with a same-day correction deadline - but it goes out.

The plan reads as authoritative because it is formatted. A clean table of workstreams and owners looks like management. It is still only as decided as the meeting was, and the [NOT DECIDED] and [UNASSIGNED] markers are the honest edges. I resist the urge to tidy them away before sharing, for the same reason my fact-checking prompts make claims visible instead of smoothing them: the marker is the work.

FAQ

Can I do this in one prompt instead of five? You can, and the output will look fine, which is the problem - the inventions are indistinguishable from the record. The chain exists to keep the model’s establishing work and its generative work in separate steps with a human between them. It costs perhaps ten extra minutes per meeting.

Does this work with an AI notetaker’s transcript instead of my notes? Yes, and often better, since nothing is lost to your writing speed. Two adjustments: tell attendees the recorder is on, and expect prompt 1’s OPEN QUESTIONS bucket to grow, because a transcript preserves every ambiguity your handwritten notes would have silently resolved.

Does this chain work in Claude as well as ChatGPT? The prompts run unchanged in either. The practical difference is where the standing context lives - custom instructions or a GPT on one side, a Claude Project with the team list and brief format as knowledge files on the other.

What if the meeting produced no real decisions? Then the workflow’s output is the discovery that it didn’t - prompt 1 returns a thin DECISIONS bucket and a long OPEN QUESTIONS list, and the honest next step is prompt 3’s follow-up message asking for those answers, not a brief built on air. I count that as the workflow succeeding, whatever it feels like.

This post is part of the prompt library. The chain pattern underneath it is prompt chaining, the brief standard it feeds is the content brief workflow, and the manager’s week it belongs to is in AI for marketing managers. The wider practice starts at the complete guide to AI marketing.