I gave a meeting-minutes tool my exact agenda words. It still lost 11 of 15 lines.
I had a theory: if you tell a meeting-minutes tool your three agenda items up front, it should be smart enough to sort any line about that topic under the right heading, regardless of how the line is phrased. I was wrong by a wide margin. Fifteen lines in, eleven of them landed in a bucket called "unassigned".
The tool is Meeting minutes draft, one of the unverified workflow candidates in the TTDU catalog. It takes 1 to 50 files (recordings transcribed on-device, or plain text notes) plus a JSON block with title, date, attendees and an agenda list, and returns a draft minutes file plus a review checklist. Nothing leaves the browser tab.
The setup
I wrote two fake meeting notes with zero real names or data: one in a timestamped transcript style ("[10:02] Speaker: new signups are up 12% this week"), the other in a handwritten-memo style ("Agenda 1: signup trend - up 12%, blog traffic suspected"). Fifteen lines total. Then I set options.agenda to the three topics both notes actually discuss: "signup trend", "cost review", "next schedule".
What I expected vs. what happened
I expected both notes to land under their matching agenda heading, since they're about the same three topics. Instead, only 4 of 15 lines got filed under an agenda section. The other 11, including a clear decision line ("Decision: keep the blog budget through next month") and a clear status update ("signups up 12% this week"), landed in "unassigned agenda".
Opening the draft text file made the pattern obvious: the handwritten-memo note used the literal phrase "Agenda 1: signup trend", and that line got filed correctly. The transcript-style note talked about the exact same signup numbers but never used the phrase "signup trend", so it got skipped entirely. This isn't semantic topic matching, it's a literal substring match against whatever agenda text you typed in.
What it got right
To be fair, the decision and action-item extraction worked flawlessly on both notes regardless of phrasing. I'd written two decision lines (one per note, worded differently) and five action-item lines, each starting with a "Decision:" or "Action:" style marker. The tool caught all seven, 2/2 and 5/5, independent of which note or format they came from.
A quirk I noticed
Even with zero audio files uploaded (text notes only), the tool still downloaded the whisper-small transcription model before running, every single time, including on a second run with the same files. It doesn't seem to affect the output, but it adds real wait time (close to 20 seconds on a cold run) for an input that never touches the transcription step.
I ran the whole thing twice with a network request listener attached. Both runs: zero requests. The output panel also carries its own warning, in plain text: "This is a draft carried over from transcripts and notes. A person needs to confirm decisions and attribution, and unconfirmed items should not be recorded as resolved." That's an honest label for what it actually does, draft assembly, not judgment.
Takeaway
If your team already writes notes using the literal agenda wording, this tool's filing will look close to magic. If your notes are freeform transcripts, the agenda sorting will miss most of the substance, even though the decision and action-item extraction stays reliable either way. I'd treat the "unassigned" bucket as mandatory reading, not an edge case.
Try it yourself at workflows.toolsthatdontupload.com, or browse the rest of the catalog at toolsthatdontupload.com. More on how these unverified candidates get tested and promoted is in the docs.