SMF·SEMBLY-AI
Can a prompt replace Sembly AI?
Meeting notes — meeting transcription and AI notes
Exhibit tracking slip
Verdict
Most meeting tools produce a document per meeting. Sembly's difference is that it produces records: an action item with an owner and a due date, a decision, a risk, each one a row you can query across every meeting you have ever had. Building that means structured extraction with a schema and validation rather than a prose summary, plus deduplication when the same task is mentioned in three consecutive meetings. That is a real fortnight, and it changes what the notes are for.
Exhibit A — The prompt
Received on31.07.2026Build a meeting notes tool that extracts typed records rather than writing a summary document.
Intake: audio or video files, normalised with ffmpeg, transcribed locally with whisper.cpp with word-level timestamps. Speakers split by an energy-and-pause heuristic and renamed by the user, propagating through the transcript. Do not claim automatic speaker identification.
Extraction, which is the point of this build. Ask the model for structured output against an explicit schema, not for prose. Four item types:
- Action: description, owner (a speaker name or an external name mentioned, or unassigned), due date if one was stated, and the transcript timestamp where it was raised.
- Decision: what was decided, who decided it, alternatives mentioned, and the timestamp.
- Risk or blocker: description, who raised it, what it blocks if stated, and the timestamp.
- Question left open: the question, who asked, and the timestamp.
Validate every extracted item against the schema and reject malformed ones rather than accepting a partially filled record. Every item must carry a timestamp and a verbatim quote from the transcript — an extracted task nobody can trace back to what was actually said is a task people will argue about. Show the quote in the interface next to the item.
Review before commit: extracted items are proposals until the user accepts them. Accept, edit or reject each one. This gate matters because extraction is confidently wrong often enough that committing automatically poisons the database.
Deduplication across meetings: when a new action closely matches an open one from a previous meeting — similar text, same owner — link them rather than creating a duplicate, and record the second mention as an update. A standing meeting otherwise produces the same six tasks every week.
The query layer, which is what this is all for. Cross-meeting views: all open actions grouped by owner; actions overdue; decisions in the last quarter, chronologically; risks raised more than once; questions never resolved. Full-text search across items and transcripts, filterable by meeting, date, speaker and item type. Each result links to the meeting and the timestamp.
Lifecycle: an action has a state (open, done, dropped) with the date and a note. Closing an action from a later meeting's extraction is one click.
Export: items as CSV, a per-meeting Markdown record, and a full JSON backup.
Use Anthropic or OpenAI, whichever key is configured. With no key, transcription, search and the item database must still work, with extraction disabled and marked.
Out of scope: joining live meetings, any bot, real-time transcription, integrations with project trackers, and team sharing.Opening prefills the prompt — press enter to run it.
Exhibit B — What you lose
- B.1 the bot that joins Zoom, Meet and Teams without being asked
- B.2 the integrations that push tasks into a project tracker
- B.3 team-wide search over everyone's meetings
- B.4 reliable automatic speaker identification
Prior art
Exhibit C — Why people still pay: capture reliability, integrations, and collaboration
Because extracted tasks only matter if they reach the place work is tracked, and the integrations that carry them there are the half of the product a local build cannot supply.
Questions
Can I import my Sembly meetings?
Transcripts and notes export as text and can be loaded as historical records. The typed items — Sembly's own structured output — do not export in a form this schema can consume, so the item database starts from the day you switch.
How accurate is structured extraction?
Good on explicit statements ("I'll send that by Friday") and poor on implied ones, which is most of a real meeting. It also invents owners when a name was mentioned nearby. The verbatim quote and the review gate exist because those two failures are frequent enough that automatic commit would make the database untrustworthy.
What does it cost to run?
Transcription is free and local. Extraction sends the transcript once and asks for structured output, so an hour-long meeting costs a few cents. A year of weekly meetings is a couple of dollars.
What is the one thing that does not survive the rebuild?
The task reaching your tracker. An action item extracted into a local database is a to-do list nobody else can see, and Sembly's value is that it lands in the tool where work actually gets assigned.
Related tools
Receipt