The AI experience I already had every day

As I described in the second post, the first version of ATLAS was a personal AI app I could access outside my MacBook.

I could keep working from my phone. Conversations were saved. It was rough, but it was a meaningful step.

Once I started using it every day, I began wanting more.

Could I make the AI experience itself a little better?

The first idea that came to mind was English.

The first layer: not a study app

What I wanted to build was not an English learning app.

I wanted to add a layer on top of the AI experience I already had.

And the first experiment for that layer was English.

Even at the time, I often thought about improving my English. But honestly, studying English felt like a chore. Opening a separate app, memorizing vocabulary, carving out review time — the process itself was a source of friction.

On the other hand, I was already using AI almost every day.

So I wondered: did I really need to keep English study separate?

If I could learn naturally inside the flow of using AI, wouldn't that be something I could keep up much longer?

My goal was simple: keep the AI experience intact, and let English learning follow naturally from it.

Prompts alone were not enough

At first I thought a few prompt tweaks would do it.

Ask the AI to suggest good English expressions after each reply, or rephrase my Korean sentences into more natural English — that seemed like it should be enough.

But once I tried it in practice, the limits showed up quickly.

English explanations started bleeding into the main answer. Responses got longer and harder to read. And most importantly, I had no control over the learning experience itself.

What I wanted was not simply "tell me some English expressions."

I wanted a specific flow: read an AI response, click on an unfamiliar word or phrase, see the meaning and explanation instantly, save it with a single click, and come back to review it later.

Language learning is usually tedious and requires deliberate time. So I wanted to reduce the cost of starting as much as possible.

Save one expression while using AI. See it again later. No separate app to open, no copy-pasting into a notebook.

A prompt could nudge the response. But it could not build that experience.

So I built the interface directly

So I built the interface I wanted.

The main chat focused on answering questions. The English layer ran quietly beside it. They were separate layers with separate roles.

The model answering questions stayed focused on the conversation. The English layer took that conversation and turned it into study material.

What I wanted was not AI teaching me English. I wanted the act of using AI to become English practice.

The English layer had two roles

The English layer did two things.

First, it analyzed what I typed.

When I typed a question, the English layer ran alongside it. If I wrote something like "나는 아직 확신이 없어," it would suggest natural English alternatives:

  • I'm not convinced yet.
  • I'm still not sure.
  • I haven't made up my mind yet.

It also surfaced common expressions or vocabulary worth knowing from that sentence — without interrupting the main answer.

Second, it analyzed the AI's response.

English was not only learnable from my own sentences. The AI's replies were full of useful expressions.

So the English layer also scanned the generated response, automatically flagged harder words and phrases, explained them, and made them saveable with a click.

The main answer was for reading. The English layer was there only when you needed it, quietly waiting beside the response.

Words and expressions were saved differently

One thing I noticed while building this: knowing a word and being able to use an expression are different skills.

A word like meticulous — once you know the meaning, you can start using it.

But a phrase like I can't quite put it into words — even if you understand it, it doesn't come to mind in real conversation.

So I split saving into two types.

Word storage keeps English and Korean together: meticulous → 꼼꼼한, subtle → 미묘한, compelling → 설득력 있는. Review means seeing the English word and recalling the meaning.

Expression storage keeps the full sentence as it appeared in context. Review means fill-in-the-blank: I can't quite ______ it into words. → put. The word is not the target — the pattern in context is.

Review mattered more than saving

Building a save feature alone was not enough.

Saved words and expressions fade without review.

So I built a simple review system alongside. Every day, saved items came up for testing. The number of correct answers and the last review time were tracked, and the review schedule followed a forgetting curve.

Words: see the English, recall the Korean. Expressions: fill in the blank in context.

Discovering something new mattered. But so did meeting it again.