The smallest process with the biggest daily payoff: from a messy transcript to owned, dated tasks — with a merge step and a check before anything reaches the task board.
Meeting Notes to Tasks — 9 nodes · 11 connections. Labels on the arrows are the conditions that take each branch.
Everyone has automated this badly at least once: transcript in, "action items" out — and half of them are vague wishes without an owner, two describe the same task in different words, and one assigns work to someone who was not even in the meeting. The failure is not the model. It is that "extract the action items" skips every decision that makes a task list usable: what counts as a task, who owns it, and what happens to duplicates.
This graph spells those decisions out. Decisions and tasks are extracted separately — a decision made in a meeting is not an action item, and mixing them is how task boards fill with noise. A labelled branch handles the tasks whose owner is unclear instead of guessing one, a merge node joins both streams before deduplication, and an evaluator checks the final list against the transcript before anything is created in your task tool. It is also the best first graph to try in nodalo: one meeting's notes, two minutes, and the output difference is visible immediately.
Each node carries a contract — what comes in, what happens, what goes out. This is the text your AI receives, node for node, when you hand the graph over.
Entry point. The raw material: a transcript, typed notes, or both — plus the attendee list with roles and the meeting date. The attendee list matters: it is the reference for resolving names like "Alex" to an actual person, and nobody outside it may end up owning a task.
Remove filler, small talk and transcription artefacts; attribute every remaining statement to a speaker from the attendee list. Do not summarise yet and do not drop anything that sounds like a commitment, however casually phrased — "I can take a look at that" survives this step. Output is a cleaned, attributed transcript.
Pull out what was decided: statements that change what is true, not what someone will do. Each decision gets one sentence in neutral wording plus who confirmed it. Open questions explicitly parked in the meeting are collected here too, marked as open — they are context for the record, never tasks.
Pull out every commitment to do something: explicit assignments and casual ones alike. For each, capture the verb phrase as stated, the likely owner as named in the conversation, and any date mentioned. Do not infer deadlines that were not said — a task without a date is honest, an invented date is a trap.
For each action item: can the owner be resolved against the attendee list without doubt? A first name that matches exactly one attendee is clear; "someone should", a name matching two people, or a person not in the meeting is not. The branches are exclusive — clear owners move on, everything else goes to clarification. No guessing, ever.
One message to the meeting organiser listing the unclear items, each with the transcript quote and a direct question: who owns this, and is it actually a task? Items confirmed with an owner rejoin the flow; items the organiser drops are recorded as dropped — silently deleting them is how "I thought you had that" happens.
Join the decision stream and the resolved task stream into one record. Merge tasks that describe the same work in different words — same owner and same outcome means same task, keep the more specific phrasing. Link each task to the decision it serves where one exists; a task serving no decision gets flagged for the check.
Check the merged list against the cleaned transcript: every task traces to a spoken commitment, every owner is on the attendee list, no date appears that was not said, and no commitment from the transcript is missing from the list. Any failure → back to merge with the specific item named, at most one round; a second failure means the notes are too ambiguous, and the list goes to the organiser marked as draft instead of being posted as fact.
The result: tasks created in the team's task tool — owner, due date where one was stated, and a link back to the meeting record with its decisions. The meeting record itself is posted to the channel where attendees will see it. Done means findable, not just written down.
A decision changes what is true ("we ship Thursday", "pricing stays"); an action item changes what someone does ("Sam updates the changelog by Wednesday"). Decisions go into the record for context, tasks go to the board. Keeping the two streams separate until the merge node is what keeps the board free of noise.
The decision node branches on exactly that: a clearly attributable owner routes the task straight to formatting, anything unclear goes to a clarify step that asks the meeting organiser — one question with the task quoted. The graph never assigns work by guessing, because a wrongly assigned task costs more trust than an unassigned one.
Any meeting that produces commitments: weeklies, project syncs, client calls, retros. For recurring meetings, keep one graph per meeting series in nodalo and hand the fresh transcript to your AI each time — the contracts stay stable, only the input changes.
Open the template, adjust the contracts, and hand the graph to your AI — as a specification it can execute, not prose it has to decode.