← Euphie Zhao
Metis · Maker · 2026

When a trial of an AI assistant ended, I rebuilt the parts I relied on — and measured mine against theirs.

Metis inbox sorting and meeting briefing, mockup
Illustrative mockup · sample data
Role
Designer and builder
Launched
September 2026, in one day
Runs on
A Mac Studio, through my Claude subscription
Reads
Gmail, Google and iCloud calendars, my knowledge base
Live
Metis home Sample AI Brief
01

The problem

I tried an AI assistant and loved two features: automatic Gmail labels, and a briefing an hour before each meeting. When the trial ended I did not want to pay for the whole product, so I built those two parts for myself.

02

Key decisions

Spend the model where it matters

Labels use the smallest model, briefings a mid-size one, and interview briefings the largest. Promotions and social mail never reach a model at all, so most email costs nothing.

Safe by default

Metis only adds labels — it never deletes, archives or moves mail. Briefings go only to me. Every script runs as a dry run first, and nothing leaves my machine except the model calls.

Knows who I’m meeting

Briefings read the people pages in my own knowledge base, so a meeting with a friend gets context, not a web search.

03

Outcome

I tested it on 126 emails the original product had already labeled. Three rounds of tuning took agreement from 67% to 83% to 89%. Briefings stay quiet too: over seven days, 25 calendar events produced 6 briefings.

It has since grown a daily work log, a morning job digest and an AI news brief, and now runs for three inboxes.

89%label agreement
100%new-sender precision, up from 12%
See it in actionA real AI Brief issue, with my product notes

Switch between English, Chinese and Korean, and hover the pins to see why each part of the email works the way it does.

Open the sample issue →