← Back to the archive

LAB ENTRY / EX—004

Context Card

A one-screen brief assembled from a calendar title, recent notes, and unresolved decisions—no transcript warehouse required.

Current statusBUILDING

AI Utility
Jun 2026

01

The question

What is the smallest useful briefing an AI tool can prepare before a meeting?

02

The hypothesis

A deliberately constrained context window will produce a more useful briefing than an exhaustive summary.

03

What I built

A local-first prototype that selects five evidence-backed items and names what it could not verify.

04

What happened

Early cards are useful when the source material is recent; stale notes remain the central problem.

05

What worked

Explicit uncertainty keeps the brief honest and makes weak context easy to spot.

06

What failed

Automatic action-item extraction overstates implied commitments, so that feature is currently disabled.

07

What surprised me

One open question is often more valuable than three paragraphs of recap.

08

What I learned

Context quality is a selection problem before it is a generation problem.

09

What happens next

Add source-age weighting and test whether users should pin one essential fact manually.