Set up your firm’s memory in Obsidian

Put the vault on the drive the office already shares, beside the project folders rather than inside them. One vault for the whole practice, because links do not cross vaults and the cross-project link is the entire point.

Your shared drive
│
├─  Projects          what you already have. leave it alone
│      2024-018 Riverside
│      2024-031 Mill Street
│
└─  Firm Memory       make this one. it is the vault

Do not point it at the whole project server. Obsidian watches every file in the tree, and a few hundred gigabytes of RVT, DWG and PDF are not notes. Make the empty folder in Explorer first, then open Obsidian and choose Open folder as vault.

The prompt does not tell the AI what folders to make. Every practice is organised differently, and a structure copied from someone else’s page rarely lasts. Instead it makes the AI interview you, and holds it to rules that stop it making things up.

This folder is an Obsidian vault. It will hold my practice's
institutional memory: what we know that is written down
nowhere, and how we decided what we decided.

Interview me before you build anything. Ask what you need to
design a structure that fits how we work, one question at a
time, and wait for each answer.

Then build it. It has to hold what happened on each project,
and the things that recur across projects as notes of their
own, so that everything touching one links to it.

Rules:
- Markdown with [[wikilinks]], never file paths.
- Invent nothing. Leave a blank and list every gap at the end.
- Everything outside this folder is read only.
- You cannot read inside RVT, DWG or scanned PDFs. Name the
  file and say so rather than guessing.

Start by asking your questions.

Point it at one finished job while it is asking, and it will fill the vault from real files rather than leave you an empty structure.

A vault is worth nothing on the day it is built and worth something on the fiftieth.

Every session starts the same way: open Cowork or ChatGPT and give it the vault folder again, exactly as you did to paste the prompt. A new chat does not remember the folder, so that is always the first move, and everything below assumes you have done it.

Every practice already has a folder per job. It answers what is in this project, and it cannot answer has this happened before. Which examiner rejected this detail, on which job, in which year. Which of the three suppliers held their dates. Those facts belong to an authority, a person or a product, not to a project. So they need to be notes that every job links to.

The same applies to the AI. A model knows architecture in general and nothing about your jobs, so it has to read them. It can only read what is written as text. A scanned markup is a picture, and what a reviewer wrote on it cannot be found. Typed into a note, the same thing can.

And when something was never written down, the model does not tell you that. It answers from general knowledge in the same confident tone it would use for your own files. So ask which file every claim came from, every time. An answer with no file behind it is an answer about buildings in general, not about yours.