Breve

Inspiration

The real cost of an important meeting is often the time spent reconstructing why it matters, what changed, where the decision lives, and what was agreed last time. Generic inbox summaries make that harder by adding more information instead of deciding what belongs to the meeting.

Breve starts from a different question: what would materially change how this person prepares for the next conversation?

What it does

Breve begins with the calendar. It checks each upcoming appointment for focused evidence: an unanswered invitation, a direct agenda or work-board link, a relevant message, a decision dependency, a scheduling conflict, or a reviewed knowledge note. Unrelated activity is filtered out.

Clicking a meeting opens a side panel with what to prepare, why the recommendation exists, current source tasks, relevant participants, focused communication, and knowledge notes. After a meeting, Breve proposes the exact knowledge destination and waits for approval before saving an additive summary.

Ask Breve supports typed and dictated requests. Read-only questions use grounded context. Requests such as drafting an email, preparing a Slack message, setting a reminder, or adding a commitment are routed to a review screen. The model never executes tools directly.

How it was built

Breve uses a Next.js and React interface with an Apple-inspired visual system. A server route optionally sends a narrow, verified context package to GPT-5.6 through the OpenAI Responses API. When no API key is configured, the same interface uses deterministic grounded answers rather than pretending an AI call succeeded.

A localhost-only Windows bridge supports optional Outlook and Obsidian actions. It uses strict origin checks, encrypted local connector storage, narrow note paths, and an explicit approval flag. Outlook email actions create drafts, not sends.

The judging repository uses a clearly labeled synthetic snapshot. The original private prototype and all connected workplace data remain outside the submission.

How Codex and GPT-5.6 were used

Codex was the primary build collaborator. It inspected an existing product handoff, preserved the strongest product idea, transformed the experience into a responsive web application, implemented the local connector bridge and approval contract, added Ask Breve and voice input, and validated the production build and visible review flows.

GPT-5.6 is the optional reasoning layer. It receives only the focused meeting snapshot and matched evidence. Its instructions prioritize tasks, decisions, owners, deadlines, response gaps, direct sources, and conflicts; distinguish missing matches from missing information; and require approval language for every proposed write.

Challenges

The hardest problem was not retrieving more information. It was enforcing relevance and trust. Breve needed to ignore unrelated mail, expose the reason behind every recommendation, avoid false execution claims, and remain useful when a connector or API key was unavailable.

A second challenge was privacy. The contest build had to preserve the complete product behavior without exporting any workplace data. The judging copy therefore uses synthetic meeting, board, message, and note examples plus portable environment-based connector configuration.

Accomplishments

• A coherent meeting-first product rather than another activity dashboard.
• Direct source tasks visible beside the meeting that needs them.
• Typed and dictated assistant requests in one calm interaction surface.
• Approval-gated email drafts, reminders, Slack-ready drafts, Ledger items, and knowledge capture.
• Honest no-key and no-connector behavior.
• A separate synthetic judging build with automated privacy scans and product-contract tests.

What was learned

AI product quality depends as much on boundaries as on intelligence. The strongest experience came from narrowing context before reasoning, making evidence inspectable, and separating suggestion from execution.

What is next

Next steps are live enterprise connector adapters, organization-defined relevance policies, stronger meeting-to-project resolution, secure shared knowledge spaces, and evaluation tools that measure whether a preparation recommendation actually changed the meeting outcome.

Built With

• codex
• gpt-5.6
• next.js
• obsidian
• openai
• outlook
• powershell
• react
• typescript
• vite