AI meeting transcribers are incredible. But the default summaries are often generic, light on action, and miss what you care about. The usual reasons: one-size-fits-all prompts and bargain models tuned for cost, not quality. So I built my own summary workflow in Lindy. It took ~15 minutes and now I get sharp briefs in Notion plus a Slack DM for anything I personally need to do.

The workflow at a glance

1. Transcript arrives
   Your transcriber drops a file or webhook with text + speaker labels.

2. Custom LLM prompt
   Run the transcript through a targeted prompt that returns structured JSON.

3. Store in Notion
   Create a page with summary, decisions, risks, and action items.

4. Personal action filter
   If any items belong to me, continue.

5. Slack DM
   Send myself a concise list of the urgent tasks.

The prompt (copy–paste)

Use a solid model and keep everything structured. Here’s the exact prompt I use:

You are the meeting scribe for our team. Analyze the transcript and return ONLY JSON.

Rules:
- Stick to facts said in the meeting. No speculation or invention.
- Keep names and nouns verbatim. Include short source quotes for key items.

JSON schema:
{
  "title": "<short meeting title>",
  "date": "<YYYY-MM-DD and timezone if present>",
  "attendees": ["<Name>"],
  "summary_bullets": ["5-8 crisp bullets with specifics (numbers, decisions, owners)"],
  "decisions": [{"what":"...", "owner":"...", "quote":"..."}],
  "risks": [{"risk":"...", "owner_or_area":"...", "quote":"..."}],
  "parking_lot": ["items deferred"],
  "actions_my": [{"task":"...", "due":"<date or empty>", "priority":"high|med|low", "quote":"..."}],
  "actions_team": [{"owner":"<Name>", "task":"...", "due":"<date or empty>", "priority":"high|med|low", "quote":"..."}]

Lindy recipe (step-by-step)

• Trigger: “New transcript file” or “Catch Hook” from your transcriber.

• Action: “Formatter → Text → Replace” to clean timestamps and filler.

• Action: “AI action” (your LLM) with the prompt above. Output is JSON.

• Action: “Notion → Create Database Item” with properties:

  • Title = title

  • Date = date

  • Attendees = multi-select from attendees

  • Summary = join of summary_bullets

  • Decisions, Risks, Parking Lot = rich text from arrays

  • Action Items (Team) = table or checklist from actions_team

  • Action Items (Me) = checklist from actions_my

• Filter: Continue only if length(actions_my) > 0.

• Action: “Slack → Send DM” to you with a compact message.

Slack DM template

*New actions for you from {{title}}*

Why this works

• Tailored structure. You tell the model exactly what to return and what you care about.

• Better models, better signal. A slightly stronger model costs pennies more but saves hours.

• Immediate accountability. Personal items land where you’ll see them.

• Searchable history. Notion becomes a clean system of record.

Tips

• Keep the transcript diarized with clear speaker names.

• Nudge accuracy by asking for short source quotes for decisions and actions.

• Add a “Follow-ups” section if your team relies on parked items.

• Log model, version, and cost per run in Notion so you can compare quality over time.