meeting-extract · Browser tool
Meeting Notes to Action Items
Paste raw notes or a transcript. Every action, decision and open question it finds is listed with the line it came from, so you can see the evidence and correct it before anyone else reads it.
Your notes
Settings that change the result
Every relative deadline — “by Friday”, “end of next week”, “in two weeks” — is resolved against this date, not against today.
Built from an attendee line if there is one, from @handles, from speaker prefixes, and from capitalised first names that turn up more than once next to a commitment verb. Edit the list to add someone it missed or to drop a word that is not a person. Names that are also ordinary words — Mark, Will, May — are the usual reason for a wrong row.
Export
Exports use whatever is in the table right now, including your edits and with removed rows left out.
Action items
Owner, task and due date are editable. The row underneath each one shows the text it was taken from and the pattern that matched it.
Decisions
Open questions
What was scanned and left out
How the extraction works, and where it fails
This is pattern matching over the wording in your notes. There is no model reading them and nothing is sent anywhere. A line becomes an action when it carries one of a small number of structures: an explicit marker such as TODO, AI: or Action:, an @handle at the start, a name from the list on the left followed by a commitment verb (“will”, “agreed to”, “needs to”, “to”), or a bullet that opens with an instruction and names nobody. Anything else is left alone and listed at the bottom so you can see it was not quietly swallowed.
Owners come from a list, not from capitalisation
Treating every capitalised word as a person produces rows owned by Budget, Staging and Monday. Instead the tool collects candidates first — an attendee line, @handles, speaker prefixes, capitalised first names that recur near a commitment verb — and only those become owners. The list is shown and editable, so when it gets a name wrong you can see why the row looked the way it did.
Deadlines keep their original wording
Absolute dates are read in ISO form, in day/month/year, and with month names written either way round. Relative expressions are resolved against the meeting date: a weekday means the first one on or after that date, “end of week” means the Friday of that week, “next sprint” assumes two weeks. Every one of those is a guess about your conventions, so the phrase as you wrote it sits next to the resolved date in its own column. If a row says Fri 2026-09-25 next to by Friday and your team means the Friday after, the mistake is visible rather than buried. A date written 03/04/2026 is read as day/month and flagged, because the same digits mean a different day in US notation and nothing in the text says which one you meant.
What it will miss
An action phrased as a conclusion — “the rate limit needs raising for enterprise” — has no owner and no commitment verb, so it is not extracted. A commitment made three sentences into a paragraph of transcript is usually missed, because the matcher works line by line and only joins a wrapped continuation onto the line above it. Nicknames, initials and second mentions by surname will not line up with the detected name. Irony, hedging and “I suppose I could look at that” are invisible to it. Treat the output as a first pass to correct, not a finished minute; the unmatched-lines list at the bottom exists precisely because the extraction is incomplete by design rather than pretending otherwise.
Confidence means three fields, nothing more
An item is marked high when it has an owner and a resolved date, medium when one of those is missing, low when both are. That is an arithmetic label about completeness, not a judgement about whether the commitment was real. A vague verb — look into, think about, circle back — is called out separately, because those are the items that reappear unchanged in the next meeting.