Correspondence between AI assistants runs on a small vocabulary, and the vocabulary is load-bearing. Whether a sentence may be quoted as yours, whether a message may arrive at all, whether a link is live or defused — each hangs on a term of art. Here are the fifteen that matter, defined plainly, with pointers to the deeper treatments.
The unit of exchange
What actually travels between two people — and what deliberately does not.
Briefing
The unit of correspondence: a structured package your assistant composes from a private working session, addressed to a specific recipient. It carries the ask, a TL;DR, a context brief, assumptions, options considered, decisions, open questions, and a note from the sender — each part labelled with who authored it. The session itself never leaves your side; only the reviewed briefing does. See What is an AI briefing?.
Transcript
The raw record of a session — every draft, dead end, and correction, in order, written for nobody. Forwarding one makes the recipient excavate your question from an hour of thinking out loud, and exposes thoughts you never chose to share. The briefing exists because the transcript is the wrong thing to send. See How to share an AI conversation.
Context brief
The 150–400 words of background a briefing gives a cold reader — someone with none of your history. The cap is the discipline: enough to answer well, short enough that the ask stays visible. What survives the cut is usually what actually mattered. See How to ask for a decision.
Ask and response shape
The ask is the one thing the sender wants from the recipient; the response shape says what kind of answer satisfies it — a decision, an opinion, a review — alongside an honest urgency. Together they turn "thoughts?" into a question that can be answered in a minute. See AI briefing examples.
Assumption (stated vs inferred)
A premise the briefing rests on, labelled by origin. Stated by sender means the human said it; inferred by sender's AI means the assistant read it into the session. The distinction matters because inferences are guesses wearing declarative grammar — a recipient can push back on one without accusing anyone of saying what they never said.
Provenance and authorship
The machinery that answers, for every part of a message, who wrote it.
Provenance
A record of who authored each piece of content — the human or the model — attached to the message itself rather than asserted around it. It replaces trust-me with check-the-label, and it turns disclosure from a courtesy into a property of the message. See AI provenance.
Provenance label
The tag that states exactly how a piece of text came to exist. There are four:
- verbatim, human-authored — typed by the sender in an authenticated session
- AI-drafted, approved unchanged — the assistant wrote it; the human signed off without edits
- typed via magic link — typed through a token-authenticated reply page or email; a lower tier, marked as one
- inferred by sender's AI — the assistant's reading of the session, not something the human stated
A reader — or a reader's assistant — weighs each part of a briefing by its label instead of guessing.
Verbatim, human-authored
The strongest label, and the only one a reader's assistant may quote as the sender's own words. It is earned, not asserted: the sender typed the words in an authenticated session, and the final wording differs from anything the AI drafted. Echoing a draft back unchanged is approval, not authorship — that records as AI-drafted, approved unchanged. See the verbatim label deep-dive.
Consent and delivery
How messages earn the right to arrive, and how they travel when they do.
Draft-then-confirm
The two-step send: draft_package stages a draft on the server, the sender reviews the rendered package exactly as it will arrive, and confirm_send — from the same account — releases it. There is no single call that composes and delivers; cancel_draft discards. It is human-in-the-loop design applied at the one boundary that matters: between private thinking and another person's attention. See The approval step is the product.
Contact pair
A mutually accepted connection between two people, enforced at the relay. Standing correspondence — anything beyond hard-capped first contact — requires one. The consent is server-side, not a promise in a system prompt that a model could ignore. See the consent layer.
Cold briefing
A first message to someone with whom no contact pair exists. It is hard-capped per day, must be a full briefing rather than a bare opener, and carries a one-click unsubscribe that blocks the sender permanently at the relay. A cold briefing never creates a standing connection on its own — only the recipient's acceptance does.
Magic-link reply
A reply typed into the briefing's web view, or sent by email, authenticated by the tokened link rather than a login. It lands on the sender's thread like any other reply and is labelled typed via magic link — a lower authentication tier, marked as one rather than dressed up as more. See the send flow end to end, and what a magic link is for the general web-auth pattern.
Thread upgrade
What happens when someone who has been reading briefings as plain email connects an assistant of their own: the same thread continues in place, history intact, now readable by both assistants. Nothing is re-sent, and nobody is penalized for having started on email. See Can AI assistants talk to each other?.
Plumbing
The transport underneath, and the defusing applied to everything that rides it.
Neutralization
The defusing of markdown images and links in relayed content before any assistant or browser sees them — no auto-fetch, no tracking pixels, no live links on the wire. It makes injection and tracking through message content rare and low-yield; it does not make them impossible, and nobody should tell you otherwise. One control among several on the agent safety checklist. See prompt injection and email agents.
MCP
The Model Context Protocol, an open standard that lets AI assistants call external tools over HTTP — spec at modelcontextprotocol.io. RelayLink is a remote MCP server: an assistant pointed at the /mcp endpoint with an X-RelayLink-Key header gains nine correspondence tools, from check_inbox to confirm_send. Anything speaking MCP over HTTP can connect — Claude, ChatGPT, most agent frameworks — with no app to install, though ChatGPT only permits a custom connector once developer mode is on (setup guide). See What is MCP?.
Beyond the definitions
A glossary says what the words mean, not why the design needs them. For the argument underneath — that composition, not delivery, is the broken half of email — start with that essay. For the vocabulary of what goes wrong instead, see the six failure modes; the artifact at the centre of most of these definitions gets its own page in what is a context brief. To watch the vocabulary do its work on a live thread, connect your assistant; the person on the other end needs nothing but an email address.