Case study
Glance
An AI email brief you can check, built by AI agents inside a design system.
Glance reads all the email since your last brief and sorts it into what needs you, what to know, and what can wait. Every fact links to its source sentence. I designed the product and its design system, and directed AI agents to build it.


The problem: I read every important email three times
I don't trust AI email summaries with anything important. So I read the summary, open the email to check it, then ask the AI what it missed. The summary saves no time.
What would a summary need to look like for me to act on it without opening the email?
This started from my own habit, not from research. I treated it as a hypothesis and tested it on three people whose email works very differently from mine.
The idea: components make it fast, sources make it trustworthy
Glance doesn't write a better paragraph. It stops writing paragraphs.
Components make it fast
A deadline becomes a time chip, a contract an amount field, a decision two option cards. You see the shape before you read a word.
Sources make it trustworthy
Tap any underlined value to see the sentence it came from.
Two kinds of content never mix
The sender's words are in quotes. Glance's own words are marked with ✦.
1. Plain AI text
What AI summaries look like today.

2. Formatted
Easier to read, still text.

3. Glance
Components, ranked, every value sourced.


Tap $48,000/yr and the source opens: who wrote it, when, and the sentence itself.
Product decisions
- A brief, not another email app
- People won't switch email apps. Replies stay in Gmail, Outlook or Yahoo Mail.
- Sorted by what you must do
- Three tabs: Needs you, Heads-up, Other mail. Each item says why it ranks where it does.
- Nothing moves while you read
- New mail waits in a “2 new” row until you add it.
- Watchlist and Memory
- Glance watches the matters you care about, learns from your moves, and follows rules you write.
- Every fix can be undone
- When Glance sorts something wrong, Move to fixes it in one step, with Undo.
Watchlist · matters Glance keeps an eye on

Memory · your rules and what it learned

The part AI didn't do
Starting over: everything passed, and it was still wrong
Version 1 passed every check. Then I used it the way Maya would, and couldn't tell what to do first.
No test catches that. The code was right; the model behind it was wrong. So I changed the model before the visuals.
Emails were sorted by type. But a promotion can matter, and a colleague can wait.
Sort by content and context: what the email says and who you are.
I had to scroll the whole page to know how much was waiting.
Three tabs with counts. The first screen answers “how much is waiting?”
“Set aside” asked me to trust mail I would never see.
Show the original email, with the words Glance used to sort it highlighted.
People read email for the deal or the school trip, not the message.
Watchlist and Memory became their own places, open to edit.
Before · version 1

After · Visual 2.0

The design system: rules the AI has to follow
The design system is code. The app and Storybook use the same tokens and components, so one change updates everything. The data names which scene card to use; the AI never picks a layout on its own.


Working with AI agents: Claude plans, Codex builds, I decide
I made every call. Claude, an AI assistant, helped me write specs, split the work and review results. Codex, an AI coding agent, wrote the code. All of it is public on GitHub.
Hand off. Rules, tokens, data, a prototype, and one issue per task.
Build. One issue, one pull request. Missing components go into the design system first.
Ask, don't invent. When the spec is silent, the agent asks and waits for my call.
Check automatically. Every pull request runs the build and 208 browser tests.
Review. I compare it with the design and the rules before I merge.
Why the steps are so small
Codex builds well, but at a fork it picks a path on its own. One wrong turn early wastes every step after it. So each task is the smallest piece I can review by eye, and a wrong turn is caught after one step, not ten.
What review caught that tests could not
“$48,000/yr” broke across two lines at 1024 px. New rule: values never break.
A whole delete row was red. Only the action should be.
The agent made an icon lighter on hover. Hover should strengthen, not fade.
The close icon should not get lighter on hover. Use --color-on-ink-muted at rest and --color-on-ink on hover and press.
Close now uses --color-on-ink-muted at rest and --color-on-ink on hover and press. Added the stories and documented the states in section 9.7.