Writing summaries that aren't bullet soup

Ask a large language model to summarize something without specific instructions, and it will produce bullets. Every time. Long bullets. Short bullets. Sub-bullets. Sometimes sub-sub-bullets.

Bullet soup.

The problem isn't bullets per se. Bullets are excellent for action items, for lists of concrete things, for step-by-step instructions. The problem is using bullets as the default format for everything, including content that has narrative structure, causal relationships, and argument.

A whiteboard from a strategy session has narrative structure. The discussion started somewhere, moved through a problem, reached a conclusion, identified next steps. That narrative doesn't fit bullets without losing the connecting tissue — the because, the therefore, the however.

### The default AI bullet problem

Here's what bullet soup looks like on a real whiteboard summary:

`` • Team discussed Q3 strategy • Pricing concerns raised • Considered expansion to enterprise • Marketing alignment needed • Budget allocation TBD • Follow-up meeting scheduled ``

This is technically accurate. It captures the topics. It tells you nothing about the relationships between them, the conclusions reached, or why any of it matters. Someone who wasn't in the meeting reads this and has no more understanding of the session than they did before reading it.

### What a good summary actually does

A good whiteboard summary functions like the smartest person in the room writing up what happened. That person:

• Connects the dots: "the team raised pricing concerns in the context of the enterprise expansion discussion"
• Captures the conclusion, not just the topic: "*decided* to delay enterprise pricing until after Q3 review"
• Notes the tension: "marketing alignment was flagged as a dependency — without it, the expansion plan can't proceed"
• Tells you what's unresolved: "budget allocation is still TBD, pending the finance review"

This reads as prose. It has structure: context, discussion, decision, open items. The reader understands what happened, not just what was mentioned.

### How we prompt for this

The BoardSnap Summary Writer agent has explicit instructions against bullet soup:

1. Write in prose, not lists. Lists are for the action items section, not the summary.
2. Use narrative connectors: because, therefore, however, as a result, given that, which means.
3. Lead with the most important conclusion, not the chronological start of the session.
4. Name the tension if there was one — what was the disagreement or constraint that shaped the discussion?
5. End with open questions, not just a restatement of action items.

The result reads differently from generic AI output. It reads more like a memo than a transcript. The goal is: someone who wasn't in the meeting reads the summary and understands not just what happened, but why it matters and what comes next.

### The action items are still a list

The action items section is deliberately a list — a structured, verb-first, owned, time-bound list. Action items are inherently list-shaped. They're discrete, parallel, individually completable. Bullets are the right format.

The distinction: narrative content gets prose, task content gets lists. Don't use the same format for both. Most AI defaults to the same format for everything, which makes both worse.

### One prompt addition that changed everything

The single prompt instruction that had the biggest impact on summary quality: "Write the summary in the past tense, from the perspective of someone who was in the room and wants the absent stakeholder to understand what happened and why."

Past tense prevents the AI from writing in the hedging present tense ("The team is considering..."). Third-person stakeholder perspective forces the summary to explain context rather than assume it.

Two constraints, dramatically better output.