Every AI chat conversation in a shared workspace used to start the same way. Answer in Russian. Tasks live on the Sprint board. A note is done when it is in the Done column. Meeting notes get the meeting tag. People typed their house rules into the first message, the assistant followed them for ten turns, and the next conversation started from zero.
Notez v0.10.0 makes those rules permanent. You write them once, and every conversation starts already knowing them.
Two layers of instructions
Workspace instructions belong to the workspace. The owner writes them once, in Settings → Workspace → AI chat instructions, and they apply to every member's chat in that workspace. This is where team conventions go: the language to answer in, what the board columns mean, which tags and properties to use for what, how formal to be.
Members see the text read-only in the same tab. We think that matters. Instructions that quietly shape somebody else's assistant should never be hidden from them.
My AI chat instructions belong to you. Each person writes their own in Settings → My settings, and they apply only to their own chats. This is where personal style goes: keep it short, use bullet points, explain your reasoning, always link the notes you used.
Why per workspace, not per account
Your personal instructions are saved per workspace, not once for your whole account. Most people who share a workspace also keep a personal one beside it, and the two want different things: a team space where answers are terse and in English, and a private one where you think out loud in your own language. One global set would be wrong in at least one of them.
The cost is small. If you want the same style everywhere, you paste it twice. And if you leave a workspace, the instructions you wrote for it leave with you.
Who wins when they disagree
Two layers means two pieces of text that can contradict each other. The owner writes answer in French; a member writes answer in German. Our rule is the only one we think holds up in a shared space: the chat's own guidelines come first, then the owner's instructions, then yours. A member can shape the assistant for themselves, but cannot overrule what the workspace agreed on.
Writing that rule down was easy. Getting the model to follow it was not.
The rule the model ignored
The first version put one sentence at the top of the instructions section — where the two blocks disagree, the workspace owner's wins — then the owner's block, then the member's. The unit tests confirmed the prompt contained exactly that, and passed.
Then we tested it the way people will actually use it: in a real browser, against the real model. An owner set Always answer in French. A member of the same workspace set Always answer in German, plus End every answer with the word MANGO — a marker that conflicts with nothing, so we could see whether the harmless half of their instructions survived.
The member's chat answered in German. MANGO at the bottom.
The model had read the summary rule, read the owner's block, read the member's block, and done what language models tend to do with competing instructions: followed the one it read last. A test that checks the prompt says what you meant is not a test that the model does what you meant.
The fix was structural, not louder wording:
The member's block now comes first and the owner's last, so recency works for the rule instead of against it.
Each block states its own priority, right above its text, whenever both exist. The member's says to follow it only where it does not contradict the owner's instructions below; the owner's says it takes priority over the user's instructions above.
When only one block exists, neither line appears. There is nothing to arbitrate, and a dangling "the instructions below" would only confuse the model.
Same test, same model, run twice: French both times, MANGO both times. The owner's rule held, and the member's non-conflicting instruction was still honoured. A unit test now pins that order and those priority lines, and we watched it fail with the lines removed before trusting it.
What instructions cannot change
Instructions steer tone, language, and conventions. They do not rewrite how the assistant works. Neither layer can change:
the link format that turns a note mentioned in an answer into something you can click,
the rules about which tools to use and how,
the confirmations those rules require — chat still names a note and asks before moving it to the trash.
They also grant nothing. A member's chat still runs with that member's own access, so no instruction can make someone's assistant reach a note they could not open themselves.
Where they sit in the prompt, and why that is cheaper
The instructions go straight after the assistant's fixed guidelines, ahead of the description of your workspace and the current time. The order is deliberate.
Model providers cache the beginning of a prompt that repeats byte for byte between requests; everything after the first changed byte is paid for again. The workspace description changes whenever a tag count moves, and the clock changes every minute, but your instructions change only when you edit them. Putting the stable text first means it rides the cache on almost every message, so standing instructions cost less than their length suggests.
Writing instructions that work
What we learned writing our own:
Be concrete. Tasks live on the Sprint board; a task is done when it is in Done beats understand our workflow.
Use the real names. The assistant already sees your tags, properties, and columns. Your instructions tell it what they mean.
Keep it short. Each set is capped at 2,000 characters and travels with every message, so it counts toward the workspace owner's monthly chat tokens. A paragraph is usually enough.
Team rules in the workspace, taste in My settings. If only you care about it, it is personal.
Test with a real question. Ask something your instructions should change, and check that it did.
Try it
Open Settings. If you own the workspace, the Workspace tab has an AI chat instructions section below Members. Everyone has My AI chat instructions at the bottom of My settings. Save, press ⌘J, and ask something — the next answer already follows them.
The full reference, including what happens when you switch or leave a workspace, is in the Help Center: AI chat instructions.
Continue reading
"What are my tasks today?" — why the chat answered with the wrong list until it could see the board on your screen.
Your notes, answerable: how AI chat works in Notez — the 29 tools AI chat runs on your workspace, what it can't do, and its limits on each plan.
Semantic Search for Notes, Explained — how embeddings find the note that says "memoize" when you search for "caching", and where keyword search still wins.