Like Talking to a Star — How We Shaped fecit’s Assistant

fecit has a conversational assistant built in. This round we worked, layer by layer, to make it feel like a calm companion sitting beside you — not a chat log. Here’s how we went about it.

Talking to a star

We turned the assistant from a chat room into a conversation with a star floating at the center.

• The star offers a line at a time, and while it’s thinking it shows a quiet … in its bubble.
• A user’s tasks aren’t spelled out in prose — they appear as tappable chips that open on tap. The chips carry the list; the assistant just opens with a light one-liner.
• When it reads back what’s inside a task, that content is set apart in a box, so “what the assistant says” and “the content” are clear at a glance.
• Instead of asking what you’d like help with, it proposes a few things you can do right now. An empty screen becomes a starting point, not a blank question.

Speaking the app’s own language

fecit gives each writing field its own name — Current, Hope, Obstacle, Stakes, Retrospect.

• We feed the assistant the user’s own label for each field so it speaks them in the exact app wording, in bold. (As a dynamic block, so the shared cross-user prompt cache stays intact.)
• Task titles are shown exactly as stored, in bold — never renamed on a whim.

Small touches, but the assistant stops sounding like it wandered in from elsewhere and starts sounding like fecit.

Fast, affordable, consistent

The assistant runs on a small, fast model (Claude Haiku) — for quick replies and a cost per message that stays comfortable.

On top of that we added small guarantees so every exchange holds the same shape:

• Lists always as chips, prose always clean.
• The brand always lowercase fecit.
• Tools always called with a valid reference — and if one’s off, it corrects course and retries right away.

The code backs up the model’s behavior one extra step. That’s what lets a fast model keep its polished feel without wobbling.

And to keep improving, we started recording per-turn token cost to the database — so we tune with data, not a hunch.

So

In a sentence: an assistant that’s fast and affordable, yet calm and consistent. Less flashy footwork, more the moment you start talking feels easy and trustworthy — that’s what we shaped, one layer at a time.

The next goal goes deeper: beyond creating tasks, toward coaching — working through a goal together. The screen and the voice we shaped today are the ground that conversation will sit on comfortably.