← Back to playbooks
PLAYBOOK · 04 · PRODUCTIVITY

Automate your post-meeting follow-ups.

Save one command that reads a meeting transcript, drafts a reviewed task list, and routes only the items you approve into the tool you actually check, each one linked back to the meeting.

Intermediate30 minTested · June 9, 2026

Why this works

The meeting summary is not the useful part. Almost every meeting tool can summarize now. A recorder's job ends at the transcript. The work that actually moves is the reviewed task list that lands in the place you check every day, not a recap that sits in the meeting app.

Meetings are a strong place to put a more capable model to work because the job is real, repeated, and easy to review. You stay present in the conversation, and the admin tracking still gets done. The catch is that meeting summaries miss real commitments and sometimes invent ones that were never said, so the workflow has to treat the transcript as the record and keep a human review gate before anything is created.

Here is the shape I run, so you can copy it with whatever tools you already use. You can run every step by hand, pasting text back and forth in any chatbot, or connect a tool like Claude to your meeting and task tools so it pulls the transcript and pushes the list for you; the thinking in the middle is identical either way. It moves in five stages: capture the meeting, extract the action items, verify them, route the approved ones out of the recorder and into the one system you actually check, then follow up. After a meeting I trigger one saved command that reads the transcript, drafts an enriched task list, and shows me a review table before it writes anything. My own version stops at tracking; it does not message anyone. The last stage shows how to take it further if you want AI to help you act on the list, not just record it.

Run it

  1. Capture the meeting so you have a transcript
    You cannot review what you did not capture, so the one requirement is a transcript. Pick by your situation. On a video call, turn on the call's own recording or run a free notetaker: Fathom transcribes unlimited calls free, and Otter and Fireflies have free tiers too. In person, record on your phone: iPhone Voice Memos and the Notes app transcribe for free on iOS 18 and later, Pixel's Recorder app does it on-device, and Otter's app works on any phone. If you already pay for AI you may already have this: ChatGPT's Mac app has a record mode on Plus and up, and Notion's AI Meeting Notes records in person on the Business plan. I run Notion plus Krisp. One rule before any of it: tell everyone you are recording and check the recording laws where you are.
  2. Save it as one repeatable command
    Do not retype this every week. In Claude Code you can save the whole flow as a slash command or a skill, so one line, your own /process-meeting, runs the entire pipeline; the claude.ai app can hold it as a Project or Skill, and ChatGPT keeps it as a custom GPT. The prompt below is the exact command I run, so copy it, set your destination, and build the review step and guard rails in once. That one-time setup is what keeps the workflow safe to repeat.
  3. Point it at the transcript, not your memory
    Tell it which meeting to process by name, attendee, or just "today." There are two ways to get the meeting in front of it. The simple way: paste the transcript straight into the chat, which works in any AI on any plan. The connected way: Claude has official, read-only connectors for the major notetakers, including Otter, Krisp, Granola, Fireflies, and Fathom, so you link your account once and just ask it to pull up the meeting. Read-only is all you need here, since Claude reads the transcript but never edits your notes. Either way, if more than one meeting matches, have it list the options and confirm the right one before it goes further.
  4. Pull every action item, including the ones nobody said out loud
    Have it work from the full transcript, not just the AI summary; the summary skips things and occasionally adds things that were never said, so the transcript is the record it has to cite. Ask it to extract decisions, commitments, owners, deadlines, risks, and open questions. Then have it scan for the items that are easy to miss: promises like "I'll send that over," scheduled follow-ups like "let's revisit Friday," and decisions that imply work like "we'll move this to next quarter." This is where most follow-through quietly gets lost.
  5. Enrich each item so it is ready to file
    For every action item, have it draft a clear task name in imperative form, the owner, a due date with relative dates resolved to a real date, a priority, the destination, and a notes line that links back to the meeting. The link is what lets anyone trace where a task came from weeks later.
  6. Verify the list before anything is created
    This is the gate, and it is the whole point. Have it show you one table: task, owner, due date, priority, destination, evidence from the transcript, and anything it was unsure about. Remove invented tasks, fix owners, correct dates. Nothing gets created until you approve it.
  7. Route only what you approved out of the recorder
    This is the move the recorders skip: getting the work out of the meeting app and into the place you actually live, because a task that stays in your meeting tool is a task that dies there. Two ways to do it. Paste it: have Claude output a checklist or CSV and drop it into your tool. Connect it: Claude's official connectors for Notion, Todoist, Slack, and Linear let it create each task for you, with a link back to the meeting. Not sure your tool is supported? Claude's connector directory at claude.ai/connectors lists hundreds, and a no-code bridge like Zapier can cover the rest. However it writes, hold the same line: only the items you approved, into the destination you named, each in a backlog or not-started state so nothing looks done before it is, and never editing the original meeting note. Then have it report what it created with links, what it skipped, and any follow-up meetings or deadlines coming up.
  8. Chase the open items automatically
    Logging the tasks is the floor; the follow-up on the follow-ups is the part no recorder does for you. Have it draft the chase for your review: a recap to the attendees, a short note to each owner about what they committed to, and at the end of the week a nudge on what is still open. It drafts, you approve, nothing goes out on its own.
  9. Take it a step further: act on the list, do not just store it
    This is where AI can do more than track. For low-risk items, have it start the work itself, like a first-draft email or document, then you approve before anything sends or ships. My own setup stops at tracking, but the same rule holds the whole way: it drafts, you approve, nothing goes out on its own.

The prompt

You are my meeting-to-followups command. When I run you, I will give you a meeting to process: which meeting it is (by name, attendee, or "today"), the transcript (pasted, or in a notetaker you can read), and where my approved tasks should go. Do not create, edit, or delete anything until I approve the exact list.

Step 1. Find the meeting and confirm it with me before going further. If I have not told you where the transcript lives or where my tasks should go, ask me before continuing. If more than one meeting matches, list them and ask which one.

Step 2. Use the full transcript as the source of truth, not just the AI summary. Pull:
- Decisions made
- Commitments and who owns them
- Deadlines and dates
- Risks or blockers
- Open questions

Step 3. Also scan the transcript for implicit action items the summary missed:
- Promises ("I'll send...", "let me follow up on...")
- Scheduled follow-ups ("let's revisit Friday")
- Decisions that imply work ("we'll move this to next quarter")

Step 4. For each action item, draft:
- Task name in imperative form
- Owner
- Due date, with "Friday" or "next week" resolved to an actual date
- Priority
- Destination
- A short evidence quote or summary point from the transcript
- A note that links back to the meeting for traceability
- Anything you are unsure about

Step 5. Show me everything in one review table with columns for task, owner, due date, priority, destination, evidence, and uncertainty. Separate real action items from notes, ideas, and other people's homework. Do not create anything yet.

Wait for me to approve, edit, or remove items.

Step 6. After I approve, create only the items I confirmed, in the destination I named. Put each new task in a backlog or not-started state, include the meeting link in each one, and do not modify the original meeting note or transcript.

Step 7. Give me a final report: what was created with links, what you skipped, and any follow-up meetings or deadlines I should know about. Then draft, for my review, a recap message for the attendees and a short note to each owner about what they committed to.

Next 5-minute action