Flyt: A Swedish writing coach I built in 10 days to learn how to design with AI agents

I'm learning Swedish, so I built the app I wanted: write a message, tap Improve, and get it back sounding more natural. Along the way I designed a process that kept four AI agents working from one set of decisions

My role
Product builder

Expertise
AI product design, Agentic workflow, Design system

Year
2026

Keep agents in sync

I built across Figma, Xcode, a Cloudflare Worker and Claude chat. By day 6, decisions made in one place weren't reaching the others, so agents filled the gaps with their own assumptions.

Lesson: collapsing environments beat syncing them. Next time, the PRD goes in the repo on day 1.

Train the model

I tested 10 real samples with a native speaker. The model gave one kind of mistake four different names and passed off two style choices as corrections. I gave it a closed list of 14 mistake types and a required-or-style flag, and a re-test confirmed the fix. Where the model was unsure, I decided: "meeting" → "möte" is a required fix.

However I have already noticed some inconsistencies in testing. Training the model takes real work. It’s hard to compete with a full team of engineers working on this task.

The human touch

In the prototype, you tapped a highlight to see why it changed. In real use, people missed that highlights could be tapped. Now the first Why card opens by itself and scrolls into view. I track whether later cards get opened by a tap, to see if the fix worked.

What’s next

Flyt is in TestFlight, and I'm learning from tester feedback and in-app events: thumbs, Why cards opened, and Copy taps. Next I'm exploring:

  • More mistake types, guided by the "other" notes in the log.

  • A mistake history, like "4th time this week you've mixed up du and dig."

  • Speech, because speaking is how you really learn a language.

Help me test Flyt. Leave your email and I'll send you an invite.