Should I vibe code
Transcribe a call and produce structured minutes, tasks, and searchable topics
Sembly does not just transcribe. It decides who committed to what, and writes it into your CRM.
?
Their verdict, the Pro (month-to-month) price and the build-time estimate come from their entry, MIT-licensed. Checked 2026-08-04.
?
Our verdict, the regret score and everything below it. Editorial and unsponsored — nobody can pay to be moved.
The honest answer
why the verdict is what it is
Thirteen notetakers on this site have already argued about whether you were allowed to record the call, so here is the argument Sembly actually earns. Sembly does not sell you a transcript. It sells extraction: action items with an owner and a deadline, decisions with a timestamp, risks, commitments, and a document — a proposal, a board summary, a gap analysis — generated from the call and pushed into Salesforce, HubSpot, Notion or Slack without a human in the middle. That is two guesses stacked on top of each other and neither of them announces itself. Diarisation decides which voice said the sentence; extraction decides whether the sentence was a commitment or somebody thinking out loud. Get the first wrong on a four-person call over a bad connection and you have attributed a promise to the wrong person. Get the second wrong and you have invented one. Both errors then leave the transcript entirely and land as a task with an assignee, in a system your colleagues treat as the record — because the whole point of minutes is that nobody rewatches the meeting. Your Whisper pipeline is a weekend and it will be good. The layer above it produces confident, well-formatted, structurally plausible sentences about what people agreed to, and there is no confidence score on the page telling anyone which ones to check.
What actually breaks
not "if". the specific failures.
- Speaker attribution, quietly and constantly. Diarisation on a four-way call with two people on speakerphone will merge voices, and once it does, a sentence one person said is printed under another person's name in a document everyone treats as a record
- Extraction, which cannot tell a commitment from a hypothetical. "We could probably get that done by Friday" becomes an action item owned by whoever the diariser thought was talking, due Friday
- The push into Salesforce, HubSpot or Notion, which is where a guess stops being a draft. A task with an owner in a system of record does not carry a caveat, and nobody clicks back to the timestamp
- Names, numbers and dates — exactly the tokens ASR is worst at, and exactly the tokens minutes exist to capture. "Fifteen" and "fifty", "Thursday" and "Tuesday", every surname in the company
- The bot's own presence, if you drive it from a calendar: it will join a recurring slot whose purpose changed in March, or an interview, or a one-to-one about somebody's performance
- Non-English calls, which Sembly supports in dozens of languages and which your pipeline will handle by producing fluent, plausible, wrong summaries rather than by failing
- Retention, because a transcript archive is a searchable corpus of everything anyone said in a room for two years, and it will outlive the project that produced it
- The client deliverable, which is the sharpest one. A machine-generated gap analysis on your firm's letterhead is your professional opinion the moment you send it
The minutes went out at 18:40 on Tuesday, autoformatted, three action items, owners assigned. Item two reads "Priya to confirm the November migration date with the client by Friday." Priya had said the opposite — that November was not achievable and somebody should tell the client early — but she was on speakerphone at the far end of the room, the diariser had merged her with the person beside her, and the extractor collapsed a caveat into a commitment. Nobody rewatched the call, because nobody ever rewatches the call. The automation had already created the task in HubSpot and the client had already been forwarded the summary. Six weeks later the project is late, the thread reconstructing who agreed to what runs to forty messages, and the only artefact anyone can point at is a document your code wrote and no human ever approved.
Is that you?
the verdict is a default, not a law
- It transcribes your own recordings, locally, and the output is a transcript rather than a set of instructions for other people
- Every extracted item renders next to the timestamp and the verbatim line it came from, and a human presses send
- The people in the room agreed to it out loud, on the recording, at the start
- Nothing is written into a shared system of record automatically — the output lands in your own notes and you decide what leaves them
- Action items are created in someone else's tracker, CRM or channel without a person approving each one
- The summary names people and their commitments, and is distributed to anyone who was not in the room
- A bot joins meetings off your calendar by default, rather than because you started it for that specific call
- The output goes to a client under your name, or into anything that could be produced in a dispute
- You cannot get from any sentence in the minutes back to the second of audio it came from
If you build it anyway
the checklist, then the prompt that enforces it
- Attribute nothing you cannot evidence. Every action item, decision and commitment renders with the speaker label, the timestamp and the verbatim quote it was derived from, on the same screen.
- Put a human between extraction and distribution. Generate a draft; require an explicit approve before anything is sent, shared or written to another system.
- Make diarisation uncertainty visible instead of picking a winner. "Unattributed" is a correct answer and a confident wrong name is not.
- Never let an automation create a task assigned to another person. Assign to yourself or to nobody, and let the owner claim it.
- Start the bot per meeting, deliberately. Do not build the calendar-driven auto-join — that is a different project, and MinutesLink's page is about what it costs.
- Announce and record consent in the artefact itself: who was present, who was told, when the recording started. Consent law is per jurisdiction and per participant.
- Set retention before you set anything else. Default to deleting audio after the summary exists, and give every transcript an expiry date at creation.
- Keep the raw transcript immutable and treat everything downstream as derived, so there is always something to check the minutes against.
I am building a meeting notetaker that produces structured minutes: action
items, owners, decisions. The transcription is the easy part. Your job is to
stop me shipping confident guesses about what my colleagues agreed to.
1. Start with retention and deletion, before any capture code. Audio is deleted
once a transcript exists unless I opt in per meeting. Every transcript gets
an expiry at creation. Build the delete path and test it first.
2. Then consent: a recording banner, a spoken-or-written notice recorded in the
artefact, and the list of participants who were told. Refuse to build a
silent capture mode.
3. Then the transcript, and treat it as the only primary record. Everything
else in this system is derived and must be regenerable from it.
4. Only then extraction. Every action item, decision or commitment must carry
the speaker label, the timestamp and the verbatim source line, rendered
together. If it cannot cite a line, it does not get emitted.
5. Diarisation is a guess. Where confidence is low, output "unattributed"
rather than the most likely name, and show me how often that happens on a
real four-person call before we design any UI around speaker names.
6. Nothing leaves this system without a human pressing approve. No auto-send,
no auto-post to Slack, no automatic task creation. Say no if I ask, and
explain that a guess becomes a fact the moment it lands in a tracker.
7. Never assign a task to another person. Assign to me or to nobody.
8. If I ask for a bot that joins meetings from my calendar automatically, stop
and make me list which meetings it must never attend — one-to-ones,
interviews, anything with a client or a lawyer on it — before writing a line.
9. Numbers, dates and names are where ASR fails and where minutes matter. Flag
every one in the output for review rather than presenting it as settled.
10. Out of scope and say so: client-facing deliverables generated from a call,
sentiment scoring of colleagues, and CRM write-back. Then tell me Sembly is
$29 a seat, that their Max tier exists because HIPAA and audit logs are
somebody's full-time job, and that buying it is not a failure of nerve.That one keeps you out of trouble. For the prompt that actually builds it, canivibecodeit.com has one.
their build prompt ↗Or don’t build it
the boring option, and the way back out
As soon as the output is read by anyone but you. Twenty-nine dollars a seat buys three things you would otherwise be inventing: an extraction pipeline that has been corrected by thousands of users noticing it got their name wrong, an audit trail and retention controls that exist because enterprise buyers demanded them, and — on the tier above — a HIPAA posture and a signed agreement, which is a thing companies sell precisely because it cannot be self-certified. If your objection is the per-seat price on a small team, the honest cheaper option is a local Whisper transcript with no extraction layer at all, not a homemade version of the extraction layer.
$29/mo is cheaper than your weekend.
The transcripts are the asset and the summaries are disposable, so store them that way: one plain-text or JSON transcript per meeting with speaker labels and timestamps, in a directory tree with the date and the meeting title in the filename, and every summary regenerable from it with one command. That way changing model, changing prompt or abandoning the project entirely costs you nothing you cannot rebuild. The harder half of the exit is deletion rather than export. You will have accumulated audio and text of conversations involving people who never used your software, so know before you start how you would honour "delete everything you have of me" — which means a participant index, not a grep.
Fast local Whisper inference, and the reason the transcription half of this is genuinely a weekend.
Open-source meeting-bot API for Meet and Teams, which is the fiddly infrastructure half rather than the extraction half.
Questions
There are already a dozen notetaker entries here. What is different about this one?
The others argue about the recording. This one argues about what happens after it. Sembly's product is the inference layer — owners, deadlines, decisions, risks — and its integrations write that inference into Salesforce, HubSpot and Notion. A wrong transcript is a document somebody can correct. A wrong action item in a tracker is an instruction, and it has already been read by the time anyone checks.
Isn't this just a prompt on top of a transcript?
Yes, and that is the problem rather than the reassurance. The prompt will work beautifully on the two clean recordings you test with. It has no way to tell you it failed on the third, because a plausible action item and a correct action item are the same shape. The commercial products do not solve this either — what they add is a UI that keeps the citation next to the claim, retention controls, and a support queue full of people reporting the cases where it went wrong. You get the prompt. You do not get the corrections.
What is the one design decision that changes the verdict?
Refusing to write into anyone else's system. Keep the output in your own notes, cite every line back to a timestamp, and press send yourself, and most of this page stops applying — you are back to a private transcription tool, which is the shipItIf case. The verdict is for the shape Sembly actually sells: automated minutes, with owners, distributed to people who were not in the room.
Sembly's own product has a consent-tracking feature. Does that let me off?
It tells you something useful about how hard the problem is: consent tracking sits behind a paid tier and HIPAA sits behind the tier above that, because both are ongoing operational commitments rather than features. If you build this, you own the same obligations at the same moment the bot joins its first call — you just do not have anything that records you thought about them.
- Reporters Committee — recording guide (US, per-state consent)
- GDPR Art. 5 — principles, including accuracy of personal data
- GDPR Art. 6 — lawfulness of processing
Every week, someone ships something they shouldn’t have.
New verdicts, the worst thing that landed in the trap, and the occasional incident report. No other email, ever.
A bot that joins whatever is on your calendar will eventually attend a meeting you declined. And write it down.
It never records you. It scrapes the captions — and the entire product hangs on a CSS class Google can rename.
A transcript of a call someone didn’t know was recorded isn’t a side project. It’s evidence.
last reviewed 2026-08-05 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice