Three meetings before lunch, or why I built my own minutes generator
My calendar does not wake up at 07:00 whispering: “One meeting today. Take it easy.”
What it usually means is three or four calls before lunch, each with its own transcript, its own cast, and—if you draw the short straw on documentation—a deadline that does not care whether you have eaten anything yet.
At that point, meeting minutes stop being “an important document.” They become a sprint. You type while replaying who said what, while praying no application module or budget figure slipped through, while hoping the output still sounds like something a director would read—not a very long WhatsApp recap wearing a tie.
Same anxiety, different Tuesday
The weekly pattern looked like this:
- Morning — cross-unit coordination, half the decisions still implicit in tone of voice.
- Midday — technical deep dive, systems and modules stacked like jargon Jenga.
- Afternoon — internal follow-up, action items breeding in the margins.
- Sometimes a fourth — “quick sync,” which in meeting dialect means an hour is quick if you squint.
After each one, the same homework: turn raw transcript into numbered, narrative, formal minutes with clean separation between discussion (Pembahasan), outcomes (Hasil Rapat), and follow-ups (Tindak Lanjut).
That is not a five-minute task. It is work you rush—and when you rush, the first casualties are technical detail, the nuance of a decision, or the one sentence that becomes next week’s actual assignment.
AI helped—until “almost right” wasn’t enough
Of course I tried the shortcut: paste the transcript into a chat model, ask for minutes.
The result was… fine. Sometimes good. Sometimes the skeleton looked right. But it kept missing in small, expensive ways:
- Voice still read like a meeting summary, not flowing institutional prose.
- Topics grouped by speaker instead of by subject matter.
- Discussion turned into wall-of-text paragraphs—while the minutes leadership actually reads are numbered and thematic.
- Action items drifted generic; names and formal verbs did not stay consistent.
- On long operational meetings, technical substance (systems, modules, infrastructure) got folded too thin.
Not because the model is clueless. Because the job context was not locked in. Every run felt like day one: re-explain format, tone, and rules—or hope this prompt lottery hits the jackpot.
Anyone who has shipped RAG in production knows that feeling: without retrieved context injected with discipline, the LLM still wanders.
From annoyance to /notulensi
So I built albanna.id/notulensi—not as an AI demo, but as a workflow tool for one job: transcript in → formal minutes out, with the rules already baked in.
Implementation-wise, it is a lightweight RAG-shaped pipeline without standing up a vector database first:
| Layer | RAG-ish role |
|---|---|
| Meeting metadata (title, date, time, venue, attendees) | Retrieved facts — hard context the model must not invent |
| Transcript / raw notes | Source document — the only substantive evidence |
| System prompt + reference examples | Policy / template — fine-tuned “how to write” for government-style minutes |
| Generator (OpenAI) | Synthesizer — narrative assembly on top of the above |
You fill the form, paste the transcript, hit generate. The model is no longer guessing format from free-form chat; it runs on a fixed track: meeting type (coordination, internal team review, technical briefing, etc.), numbered discussion sections, separated outcomes and follow-ups, formal narrative voice, and a hard rule not to fabricate beyond the transcript.
That is the gap versus one-shot chat: retrieval is disciplined—metadata and transcript always ship together; generation is constrained—a long system prompt plus reference minutes written for directors, not for demos.
What I got back
The AI is not perfect. Garbage transcripts still produce garbage minutes. You still read everything before it goes upstream.
What returned to my calendar was time spent re-teaching format and energy spent rebuilding structure from scratch. A four-meeting day feels like four drafts to edit, not four mountains to quarry.
If you write minutes on a deadline and have been annoyed by AI that is “close” but not quite formal enough: the tool lives at albanna.id/notulensi. It is intentionally not in the blog navbar—workbench, not billboard.
I still sign off on the final text before it reaches leadership. That step stays human. I only wanted to automate the part that is boring, repetitive, and easy to get wrong: reconstructing a format the machine should already know.