Where We Use AI

I wanted to add AI. There were things that were tedious for people to do, or things that took too long, and AI seemed like a perfect fit.

Here’s where we’re using it and why. Some of it works well. Some of it is still rough.

Template Recommendations

When you start typing a title for a new task, similar templates appear below. “You made something like this before.”

Calling this AI feels generous. It’s really closer to search — finding existing templates by title and showing them. Frequently used ones float to the top, pinned ones get priority. Up to three suggestions.

There’s a 300-millisecond debounce. Without it, the list flickers with every keystroke, which is distracting.

It’s a simple feature, but it’s the one I use the most. If you repeat similar tasks, it’s surprisingly convenient. Type a few letters and the template you refined last time is right there.

There’s also ghost text. As you type, the recommended template’s name fills in faintly. Like it if you want — one tap to accept. Don’t like it — just keep typing.

Auto-generating Sub-tasks

This was a “wouldn’t it be nice” feature.

AI looks at the title and description of a template or task, then generates sub-tasks. Type “Write a blog post” and it suggests “Pick a topic,” “Write draft,” “Edit,” “Publish.”

We use OpenAI’s GPT model for this. There are separate prompts for Korean and English, with a JSON schema for structured output. It generates 3 to 5 sub-tasks and automatically creates dependency chains between them.

It’s not perfect. Sometimes the suggestions are too generic, sometimes they miss the context entirely. But having something is better than staring at a blank screen. You can always fix it from there.

There’s a loading animation with spinning stars. Generation takes a few seconds, and leaving the screen blank during that time felt anxious. A small detail, but these things affect how it feels.

There’s a daily usage limit. A compromise due to cost. I wanted it to be unlimited, but that wasn’t realistic.

Improve Suggestions

When you complete a task, if it differs from the original template, you get a question: “This is different from the template. Want to apply the changes?”

Say you created a task from your “Weekly Report” template, then modified the description and added sub-tasks along the way. After completion, it suggests applying those changes back to the template.

This matters because templates aren’t meant to be created once and forgotten. They should get a little better each time. But after finishing a task, going back to edit the template is a chore. This feature reduces that friction.

One tap to apply. Ignore it if you don’t care. You can turn off the suggestion entirely in template settings.

Overview Recommendations

When browsing other people’s overviews in the community, AI recommends overviews based on category and language.

Technically, we generate embeddings with OpenAI’s text-embedding-3-small model and use MongoDB Atlas Vector Search to find similar content. User interests are also maintained as embeddings — updated incrementally whenever they fill out a survey or create tasks. To keep recommendations from getting stale, we mix in a bit of Gaussian noise so results vary slightly each time.

Still refining this. Recommendation quality should improve as more content accumulates.

The spinning star animation appears here too. It shows for at least 2.5 seconds. If it finishes too quickly, it feels like it didn’t really try. That’s a bit of a trick, and I’m slightly embarrassed about it.

Save and Focus Suggestions

If you create tasks with similar names repeatedly, it asks: “You make this often. Want to save it as a template?”

When you start a task, it asks: “Focus on this?” Focusing pins it to the top of the list.

Both are closer to pattern detection than AI. But asking the right question at the right moment turned out to be harder than expected. Too often and it’s annoying. Too rarely and it’s invisible.

Still finding that balance.

Being Honest

The AI features are still being refined, but I use them every day and they’re getting a little better each time. Cost is a factor, but they’re solid enough to be useful.

But the direction is clear. AI isn’t here to do your tasks for you. It’s here to make starting and finishing a little smoother. When you’re staring at a blank screen not knowing what to write, it throws out a line. When you’ve just finished and can’t be bothered to organize what you learned, it asks once.

It’s not grand AI. It’s more like a nudge from the side. I think that’s enough, and doing that “enough” well turns out to be surprisingly hard.

Still building. If something feels off, let me know.

---

AI is a tool. A good tool makes work a little easier.