How Briefi briefs work — what the AI reads, what it writes

Most 'AI meeting brief' tools summarise an email thread. That's not a brief. A brief is what you'd want a smart chief of staff to write the night before — who's in the room, why they care, what's likely to come up, what to ask, and what's a red flag. Restating an email thread isn't that.

This post is the long-form explanation of what Briefi reads when you hit 'Generate brief', what it does with each input, and what the output is shaped to look like.

What goes into a brief

Five inputs, in priority order. The calendar event itself — title, attendees, location, conferencing link, any description text. Prior interactions with the attendees — every previous meeting we have in your synced calendar, every recent thread from the attendees in your inbox, the contact record if we have one. The discovery guide selected for this meeting (Briefi auto-picks via the attendee + contact data, overrideable by the user). Optional operator context — a free-text field on the Prepare page where you type or dictate background the AI should treat as authoritative. Optional web research — when the toggle is on, Briefi pulls public background on the primary external attendee and their company before drafting.

Every input is decrypted server-side before the model sees it. Nothing about your meeting data is sent to the AI provider unencrypted in transit, but inside the request the prompt itself does include the relevant content. We don't pretend otherwise. Our subprocessor list on the security page names the providers in play.

How the prompt is shaped

The system prompt establishes the brief's structure: a one-paragraph context section, three smart questions, two red flags to watch for, a recommended outcome. The user prompt is the inputs, concatenated and labelled. Voice profile is injected last as a style hint — 'write in this user's tone, mirror when their samples are well-written, default to high-quality prose otherwise'. Output language matches the user's working language; AU English is the AU-default but the prompt explicitly tells the model to switch when the input language is different.

Briefi does the writing. The brief costs 10 credits, charged when it lands successfully. If the generation fails or returns nothing, you don't get charged.

What comes out

A worked example. Meeting: 'Discovery — Pricing review with Acme Co'. Attendees: a head of operations from a 200-person SaaS company. Prior interactions: one introductory call last week, two follow-up emails.

The output reads something like: 'Context: Acme has been on competitor X for three years, recently restructured the operations team, and the head-of-ops is new in role. The intro call surfaced a budget conversation that wasn't closed. Three smart questions: (1) what changed in the restructure that made the current setup intolerable; (2) where in the year is the budget cycle; (3) who else is in the decision room you haven't named. Red flags: head-of-ops mentioned a 'quick win in Q1' — projects in their space rarely deliver in under 90 days; competitor X is unlikely to lose this account without contract leverage. Recommended outcome: leave with the names of the other two decision-makers and a soft-yes on a proposal in 7 days.'

That's the shape every brief takes. Adapted per meeting type via the discovery guide, but the spine is consistent.

The web research toggle

Opt-in on the Prepare page, gated by the `web_research` feature flag (Solo plan and above). When on, Briefi attaches a web-search capability to the brief generator with up to three queries per run, and instructs it to gather public background on the primary external attendee + their company before drafting. Output reflects what it found — recent press, headcount changes, funding events, product announcements. The cost is folded into the existing 10-credit brief charge; no separate billing path.

Free plan sees the toggle as an upsell discoverability cue and the server returns 403 on submit. Same pattern as every Pro+ feature on the product.

Why a brief is one shot, not a chat

Briefs are a single generation, not a conversation. You don't 'chat' with the brief. The decision was deliberate — chat tools train users to expect dialogue, which is the wrong shape for pre-meeting prep. You read the brief, you walk in, you run the meeting. If the brief was wrong, you regenerate (1 credit, not 10) with refined context.

Chat works for the MCP server, where the agent is doing multi-step work across meetings. For the brief itself, one-shot is the right register.

What it doesn't do

Briefs don't pretend to know things they can't. If your inbox has no history with the attendee, the brief says so. If the calendar event has no description and no prior context, the brief still produces a structure but flags the limited input. The worst failure mode in meeting AI is confident hallucination; we work harder than usual to avoid it.

Briefs also don't try to script the call. The discovery guide does the scripting; the brief is the operator's context layer above the script.

The product page walks through Prepare in screenshots. For the script side of the loop, see why discovery guides beat checklists. And for the founder's view of why a brief is the first thing every meeting needs, see why I built Briefi.