Your notes carry meaning that isn't on the page. What a project is, who a client is, which tag means what, how a page should be read. Steering is how you hand that context to Thoute's AI once, so every AI feature reasons the way you would — instead of guessing from the words in front of it.
Four layers of context, from broadest to most specific, all optional and all end-to-end encrypted:
| Layer | Scope | Set it in |
|---|---|---|
| Vault brief | The whole vault, always | Settings → Vault brief |
| Entity glossary | Named people, projects, clients | Settings → Glossary |
| Tag context | Everything carrying a tag | The tag's editor |
| Page & subtree context | One page or one branch of the outline | The page or the bullet's ••• menu |
Every layer is guidance, not a source. Thoute still answers from and cites your actual notes; the context you author just tells the model how to read them. And all of it is authored by you in plain language, stored encrypted, and synced across your devices — the server never sees any of it in the clear.
Steering feeds every place Thoute reasons over your notes:
You don't invoke steering. It's applied automatically whenever one of those touches content that a layer covers.
The vault brief is one standing note about you and this vault — the "read me first" every AI feature sees. Who you are, what the vault is for, the projects you're in the middle of, the tone you want answers in, and any always/never rules.
Write it once in Settings → Vault brief. It's injected as ambient grounding into every AI run, so you never re-explain your workspace.
Good things to put in a brief:
Keep it tight. The brief is grounding, not an encyclopedia — a few short paragraphs beats a wall of text.
Tags carry a #. Entities are the plain proper nouns that fill your notes and mean nothing to a model on their own — Acme, Priya, Project Halo. The glossary is where you define them once so AI knows the real-world thing behind the name.
Define them in Settings → Glossary (Pro plan). Each entry has:
Acme.Acme Corp, ACME.When Ask AI, a recall, or an automation touches content that mentions an entity — by name or by any alias — that entity's definition is quietly handed to the model. So a question like "what's the status of my biggest client?" connects to Acme even though you never typed "Acme," because the brief and the glossary together tell the model who your biggest client is.
Matching is whole-word and case-insensitive. Very short names (one or two letters) are skipped — they'd fire on ordinary words — so give an entity a distinctive name or alias if you want it recognized.
A tag can carry a short description of what it means and how the AI should read it. This lives in the tag's own editor, not in Settings — see Live tags → AI context for the full walkthrough.
In short: open a tag, expand AI context, and fill in a description, an optional interpretation prompt, and any related tags. When you ask a question that names that tag, its meaning rides along. Authoring tag context is a Pro feature.
Sometimes meaning is structural — "this page is my client roster," "this branch is a research log, treat each child as a source." You can attach an interpretation prompt (and metadata) to a single page or a single branch of the outline.
••• menu and choose Add AI context…. It applies to that bullet and everything beneath it.Both open a small AI context panel with a freeform prompt and optional metadata; changes save automatically. Whenever an AI run draws on content from that page or branch, the prompt is included so the structure is read the way you intended.
Because a branch's context covers everything under it, you can set it high on a subtree and let it apply to the whole thing — the nearest ancestor with context wins.
Every steering layer is stored exactly like your notes: client-side encrypted, zero-knowledge. The server holds only ciphertext and never sees your brief, entities, tag context, or page prompts in the clear.
The one time any of it leaves your device is the same moment your notes do: when you run an AI feature with Cloud AI enabled for the vault (off by default). At that point the relevant context is sent, alongside your sources, through Thoute's server-side proxy to Google's Gemini API to answer the request. The proxy doesn't store the plaintext, and anything written back to your vault is encrypted again. See Privacy for the full model.