Meeting notes extraction schema with decisions, owners and quotes
A strict JSON Schema plus extraction prompt that turns a meeting transcript into decisions, action items with owners and due dates, open questions and risks, each backed by a verbatim quote so readers can verify it.
File in Operations by AISkillHub Library
What it does
AI meeting summaries go wrong in predictable ways: proposals reported as decisions, tasks assigned to someone who never agreed, due dates that were never said. This schema and prompt are built to prevent those errors and make the remaining ones easy to spot.
**What the schema captures**
- `meeting_title`, `meeting_date` (null unless stated), `attendees` (only people who spoke or were listed as present) - `summary`: three to five sentences, no action items - `decisions`: decision, who made it (or null for group consensus), and a **verbatim quote** - `action_items`: task starting with a verb, owner (null if nobody took it), due date as stated, and a **verbatim quote** - `open_questions` and `risks`: only those raised in the meeting
**Why quotes**
Requiring a short verbatim excerpt for every decision and action item does two things. It discourages the model from inventing items, because each needs supporting text, and it lets a reader check any item in seconds by searching the transcript. If a quote is not in the transcript, treat the item as suspect.
**Files**
- `schemas/meeting-notes.schema.json`: draft 2020-12 schema, every property required, `additionalProperties: false` on every object, nullable fields typed as `["string", "null"]`. Compatible with strict structured output modes that require this style. - `schemas/meeting-notes.prompt.md`: the extraction instructions, with one variable, `{{transcript}}`.
**What it will not do**
- Transcribe audio. Feed it text from your transcription tool. - Resolve who "he" or "the team" refers to when the transcript does not say. Those owners stay null by design. - Track action items across meetings. Store the JSON and compare runs yourself if you need that.
Very long transcripts may exceed the model's context window; split by agenda item and merge results if needed.
How to use it
1. Copy `schemas/meeting-notes.schema.json` and `schemas/meeting-notes.prompt.md` into your project. 2. Fill `{{transcript}}` in the prompt with the full transcript text, including speaker names if your transcription tool provides them. 3. Send the prompt with the schema: - OpenAI: use the schema as a strict JSON schema response format. - Claude: define a tool such as `record_meeting_notes` with the schema as `input_schema` and force it with `tool_choice`. 4. Validate the result against the schema in code. 5. Optional quality check in code: confirm every `quote` value appears in the transcript (after normalizing whitespace) and flag items whose quote does not. 6. Post the decisions and action items to your tracker, and keep null owners visible so someone assigns them.
Files
Save each file at the path shown, relative to your project root.