Lumen

Lumen began with a complaint I kept hearing from writers: AI tools are impressive in a demo and exhausting in practice. Every suggestion arrives as a finished paragraph. Every generated sentence sounds like the average of the internet. The tool accelerates exactly the part of writing that was never slow.

Designing for the pause

The first thing we removed was the chat box. Instead of a conversation you have to maintain, Lumen reads the document as it exists — outline, argument, evidence, tone — and offers suggestions in the margin, one clause at a time, never inside the sentence you are typing. Accepting one takes a keystroke. Ignoring one takes the same keystroke, in the other direction.

That constraint did more for the product than any model upgrade. When suggestions became cheap to refuse, writers stopped bracing for them. The assistant stopped feeling like a second author and started feeling like a good editor sitting across the table.

The quiet parts

Most of the design work was not the AI at all. It was text measurement, the way the margin collapses on a narrow window, and an undo stack that has to feel like one continuous gesture across human edits and machine edits. We rebuilt undo four times before it stopped surprising people.

The best compliment a writing tool can receive is that you forget it is there.

We also made a stubborn typographic bet: one serif, one sans, both variable, no decoration. A writing surface should be furniture. Every weight is tuned by hand, and the measured column is capped at roughly 68 characters regardless of window width.

What the beta taught me

• Writers used the assistant more when accepting a suggestion was harder, not easier — deliberation kept trust high.
• The library mattered as much as the editor. Search that remembers intent beats search that matches text.
• A local-first file format bought us patience: people could try Lumen without moving their whole practice into it.

What I would do differently

I would instrument the first five minutes far earlier. Our best retention signal turned out to be a mundane one — whether a person imported a document in their first session. We discovered it in week nine, from support threads, instead of week one, from telemetry.