Context engine (client brain)

What a client profile contains

A client profile is keyed to a domain (e.g. acmecorp.com) and auto-loads whenever you open a matching thread. Every field you fill in becomes something the agent reads while it drafts. Profiles hold: status, point of contact, tier, tags, open loops (action items), 'don't forget' notes, guardrails, and typed variables.

Variables resolve into drafts

Variables are typed {key: value} pairs. When the agent drafts, it resolves each variable into its real value — never a placeholder or a guess. If a variable has no value, the agent leaves it unfilled and flags it rather than inventing one.

Variables live at two scopes: global (available everywhere) and per-client (defined on a specific profile). A per-client value takes over from the global one for that client, so you can keep a general default and override it where a particular account differs.

Guardrails and don't-forget notes

Guardrails are must-follow rules attached to a profile — for example 'never mention pricing without approval' or 'always CC the account manager'. Don't-forget notes are reminders you want kept in mind on every reply.

Both are injected straight into the agent's prompt — for interactive drafts you write with Copilot and for anything Autopilot sends on its own. That means the same rules apply whether a human is in the loop or not.

Context folders

Context folders are shared reference notes attached to a mailbox. A folder holds short text or markdown items that steer the agent's voice and give it background facts — house style, product details, standard policies — so its drafts sound like you and stay accurate across every thread on that mailbox.

Use a folder for anything that isn't specific to one client: the way you like to sign off, facts about your own company, or a policy that applies broadly. Client profiles handle the per-account specifics on top of this shared layer.

Uploaded reference files

You can attach reference files — PDF, DOCX, or Markdown — to a profile or to a context folder. On upload they're converted to markdown, and the agent reads them while drafting.

This is how you give the agent longer-form source material: a contract, a brand guide, a spec sheet, a services one-pager. Attach it once and the agent can draw on it whenever it writes for that client or on that mailbox.

Setting it up

A typical setup goes profile first, then the shared layers around it:

1. 1

   Create a client profile

   Add a profile keyed to the client's domain and fill in the fields you care about — status, point of contact, tier, tags, open loops, and don't-forget notes. It will auto-load whenever you open a thread with that client.

2. 2

   Add variables and guardrails

   Define typed {key: value} variables (global or per-client) so real values drop into drafts, and add guardrails — must-follow rules like 'never mention pricing without approval' — that the agent obeys on every draft and autonomous send.

3. 3

   Attach a context folder

   Add a context folder to the mailbox with shared reference notes — your house voice and general facts — as short text or markdown items. The agent reads these to keep its drafts on-voice and accurate across every thread.

4. 4

   Upload a reference file

   Attach a PDF, DOCX, or Markdown file to the profile or folder. It's converted to markdown on upload, and the agent reads it while drafting — handy for contracts, brand guides, or spec sheets.