SMF·SUPERNORMAL
Can a prompt replace Supernormal?
Meeting notes — meeting transcription and AI notes
Exhibit tracking slip
Verdict
Supernormal's useful idea is templates: a customer call always yields the same six sections, a standup always yields the same three, and consistency is what makes a year of notes searchable rather than a pile of prose. Building that means per-section extraction instructions rather than one summary prompt, plus a validation step so a section with nothing to say says so instead of inventing content. That is genuinely a fortnight, and the bot that joins calls is out of reach regardless.
Exhibit A — The prompt
Received on31.07.2026Build a meeting-notes tool where the output shape is defined by a user-authored template.
Intake: audio or video files normalised with ffmpeg, transcribed locally with whisper.cpp with word-level timestamps. Speaker splitting by an energy-and-pause heuristic, renameable by the user with the change propagating. Do not claim automatic speaker identification.
Templates, which are the point of this build. A template is a named, ordered list of sections. Each section carries:
- A heading.
- An extraction instruction written by the user, in their own words, describing what belongs in this section.
- An output shape: prose, bullet list, or a table with named columns.
- A required flag.
- An optional example, which does more to pin down the result than any amount of instruction.
Generation runs per section, not as one call for the whole document. Each section's request receives the transcript and only that section's instruction, shape and example. This costs more calls and produces markedly better results than asking for a whole structured document at once, because each request has one job.
Emptiness handled honestly. Instruct each section to return an explicit empty marker when the transcript contains nothing relevant, and render that as "nothing recorded for this section" rather than dropping the heading or, worse, letting the model fill it with something plausible. A template that always produces content is a template that produces fiction, and this is the single most important rule in the build.
Grounding: every bullet or row carries the transcript timestamp it came from, clickable to play that moment. A section rendered as prose carries timestamps at paragraph level. Extraction that cannot be checked against what was said is extraction nobody should trust.
Template management: templates stored as files so a set can be shared or versioned. Editing a template creates a new version; every generated note records which version produced it, so a change in output can be traced to a change in instruction. Ship no templates — the first run walks the user through building one from a real meeting they already have.
Also build: full-text search across notes and transcripts, filters by template and date, export as Markdown and JSON, and a re-run action that regenerates a note under a newer template version while keeping the original.
Use Anthropic or OpenAI, whichever key is configured. With no key, transcription, search and template editing must still work.
Out of scope: joining live meetings, any bot, real-time transcription, integrations with other systems, 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 calls automatically
- B.2 the integrations that file notes into a CRM or a document store
- B.3 shared team templates and a common note standard
- B.4 mobile capture and cross-device sync
Prior art
Exhibit C — Why people still pay: capture reliability, integrations, and collaboration
Because a note standard only works if a team shares it, and shared templates plus automatic capture are what make everyone's notes comparable. One person with a good template is a habit, not a standard.
Questions
Can I import my Supernormal notes and templates?
Notes export as text and load as historical records. Templates do not export in a structured form, so each is rewritten — which takes minutes and is worth doing carefully, since the per-section instructions are where the output quality lives.
Why generate section by section?
Because one call asked to fill six differently shaped sections does all of them adequately and none of them well. Separate calls with one instruction each produce noticeably better sections, and they make it possible to regenerate one section without disturbing the rest.
What does it cost to run?
Transcription is free and local. Six sections against an hour-long transcript is six calls each sending the transcript, so perhaps ten to twenty cents per meeting — the most expensive of the meeting-notes builds in this catalogue, and the trade for a better result.
What is the one thing that does not survive the rebuild?
Being in the room. Supernormal's bot joins the call whether you remembered or not, and a template only helps for meetings that got captured. This build sees the files you deliberately recorded, which is a smaller set than you think.
Related tools
Receipt