Meeting Summary — Vague Prompt vs Structured Prompt

Hagar got asked to send a meeting summary to her team after a planning call about Q3 numbers. She tried two prompts. Same transcript. Same model. Two very different outputs. The lesson is the same as Module 1 — but on a more realistic, work-shaped task.

The transcript both prompts use

Sarah: Q3 numbers came in 8% under target. The new pricing tiers we shipped
in July are not converting like we hoped.
Tom: I think it is the messaging on the landing page, not the price.
Sarah: Can we A/B test new copy by next Friday? I will own the experiment design.
Tom: Yes. I will pull the analytics. We need at least 2000 visitors per variant.
Sarah: Agreed. Let's also revisit the discount tier — 20% off feels too steep.

Attempt 1 — vague prompt

Prompt:

Summarise this meeting:

[transcript]

Captured output:

# Meeting Summary

**Main Issue:** Q3 revenue is 8% below target due to poor conversion on new pricing tiers launched in July.

**Key Discussion:**
- Tom believes the problem is landing page messaging rather than pricing itself
- Sarah suspects the 20% discount tier may be too aggressive

**Action Items:**
- **Sarah:** Design A/B test for new landing page copy (deadline: next Friday)
- **Tom:** Pull analytics data (minimum 2,000 visitors needed per variant)
- **Both:** Review the 20% discount tier

Captured from Claude Sonnet 4.5 (claude-sonnet-4-5) on 2026-04-27. Re-runs may differ slightly.

This is genuinely good. The model picked sensible buckets (Main Issue, Key Discussion, Action Items) and bolded names. Hagar could ship this. But notice — the categories are the model's choice, not Hagar's. If her team has a different convention, she has to edit.

Attempt 2 — structured prompt

Summarise this meeting using exactly this structure:

DECISIONS:
- (each on one line)

OPEN QUESTIONS:
- (each on one line)

ACTION ITEMS:
- [Owner] [Action] (due date if mentioned)

Transcript:
[transcript]

Captured output:

DECISIONS:
- Run an A/B test on landing page copy by next Friday
- Need at least 2000 visitors per variant for the test

OPEN QUESTIONS:
- Is the issue with pricing tiers or landing page messaging?
- Should the 20% discount tier be adjusted?

ACTION ITEMS:
- [Sarah] Design A/B test experiment (due next Friday)
- [Tom] Pull analytics for the test (due next Friday)

Captured from Claude Sonnet 4.5 (claude-sonnet-4-5) on 2026-04-27. Re-runs may differ slightly.

Different shape entirely. The model used Hagar's categories (decisions, open questions, action items), respected the bullet format she specified, and used the [Owner] [Action] syntax for action items.

When each one is right

Use this	When
Vague prompt	One-off, you don't care about format, you'll skim and copy-paste pieces.
Structured prompt	You're writing into a template, your team has a convention, or this summary will be machine-read (issues tracker, CRM).

The structured version is the one that compounds. Once Hagar has it, she can use it on every meeting. She doesn't have to rewrite the categories each time.

The "open questions" trick

Notice the structured prompt has an OPEN QUESTIONS: section. The model put two real ones in there: "Is the issue with pricing tiers or landing page messaging?" and "Should the 20% discount tier be adjusted?" Both are unresolved threads from the actual conversation.

A vague summary smooths these over. A structured summary surfaces them — which is exactly what you want, because unresolved questions are the most likely things to fall through the cracks.

If your meeting summaries never feel useful, this section alone is worth adopting. Most meetings have at least one open question; most meeting notes don't capture it.

The takeaway

Same model, same input, two prompts, two utility levels. The structured prompt is more typing up front and saves more time downstream. By the third meeting, the structured prompt is muscle memory.

Next: iterating on a LinkedIn rewrite — what to do when the first attempt is generic and overflowing. :::