Ten opinions, six ways of working

Third week of the semester, and the same explanation of Bayes' theorem has been asked for four times — in four different chats, all of them gone. With SumizAI the fourth time never happens. The good explanation arrives, it gets saved, and it is a file called bayes-theorem.md in the course folder from that moment on.

By the end of term the folder is a set of notes written in the exact language that made the concept land the first time — not a textbook's language, the student's own. Revision stops being re-reading and becomes re-opening.

Four hundred papers in, the problem is never finding a source — it is remembering that you already read it and what you thought at the time. Each paper gets summarised in a conversation, each summary becomes a note, and the notes link to each other with ordinary Markdown links, so a claim in one note points at the paper that supports it.

The duplicate check earns its keep here. Two papers making the same argument produce two near-identical notes, SumizAI flags the overlap, and one merged note survives with both citations instead of two half-notes drifting apart.

Documentation written after the fact is documentation nobody writes. Here it is a by-product: the debugging conversation that finally explained why the connection pool was exhausting itself is the note, saved in the repo's notes folder as a plain .md file that git tracks like any other.

Because the vault is just files on disk, it travels with the project. Clone the repo six months later and the reasoning behind the odd retry logic is sitting next to the code that does it, in the same commit history.

Five clients, five domains, five vocabularies. Five vaults keep them apart — separate folders, separate tables of contents, no chance of a retailer's pricing model leaking into a hospital deck.

When a proposal is due, the thinking is already written down. The relevant notes go to PowerPoint, the appendix goes to PDF, and the deck that used to take a Sunday takes an afternoon — because the hard part, the reasoning, happened weeks ago in a chat window and never evaporated.

Every essay starts as a question asked out loud. The good half of the answer becomes a note; the note joins twenty others that were already circling the same theme, and the table of contents quietly shows the shape of a book that nobody sat down to plan.

Nothing is locked in a proprietary editor. The notes are Markdown, so they open in Obsidian, in a text editor, in whatever comes next — the archive outlives the tool that made it.

Pricing, hiring, tax residency, a database migration — four decisions in one week, each researched in a chat, each one gone by Friday unless something catches it. SumizAI catches it, and the vault becomes the company's memory: what was decided, what the alternatives were, and why the cheaper option lost.

The bill is a dollar a month on your own AI key, which for a one-person business is the kind of decision that takes four seconds. Stop paying and the notes still open and still export — only new conversations stop.