Prompt engineering for meeting summaries

Effective meeting summaries require precise prompt constraints. If prompt instructions are too loose, models can generate verbose prose that obscures key decisions. Oyma allows you to fine-tune system prompts, temperature parameters, and output schemas.

The anatomy of an executive summary prompt

A high-performing meeting synthesis prompt consists of four structural layers:

1. Role Definition: Instructs the model on its persona (e.g. “You are an executive chief of staff specializing in software engineering governance”).
2. Context Grounding: Passes the raw transcript and participant roster.
3. Negative Constraints: Explicitly forbids speculation, ungrounded action items, or conversational filler.
4. Output Schema: Mandates strict CommonMark formatting with headings, tables, and task checklists.

Recommended custom prompt template

Paste this template into Settings › AI › Custom Summary Prompt:

You are an expert technical note-taker. Summarize the following meeting transcript accurately and concisely.

Rules:
1. ONLY include facts explicitly stated in the transcript. Do NOT extrapolate or assume.
2. Structure output using strict CommonMark Markdown.
3. Attribute decisions and action items to named participants whenever possible.

Output Format:
## Executive Summary
[3 bullet points summarizing core outcomes]

## Key Technical Decisions
| Decision | Rationale | Alternatives Rejected |
|---|---|---|

## Action Items
- [ ] [Owner]: [Task description] ([Deadline if mentioned])

## Open Blockers & Follow-Ups
- [List unresolved questions]

Tuning model parameters

• Temperature (0.1–0.3): Keeps the model grounded strictly in transcript text. Avoid temperatures above 0.5 for meeting documentation.
• Top-P (0.9): Limits token selection to high-probability linguistic paths.
• Max Tokens (2,048): Generates comprehensive multi-section records without premature truncation.

Related guides

• Summary archetypes in AI meeting summary archetypes.
• Ollama setup in Ollama local setup and model configuration.
• Token controls in Token usage and cost controls.