Now the Star Asks Back

In the last post, I took the guide out of its mode. Focus an empty field and that field’s question surfaces in the top bar. “What are you hoping for?”, “What’s standing in the way?” — the star asks, you fill it in.

But that question was the same for everyone. Open the expectation field and the same line appears, whoever you are. It helps, yet in the moment you’re most stuck — you’ve written only “evening meds” and the cursor blinks in the next field — a generic question is often one more thing to be stuck on.

So I gave the star the ability to read what you actually wrote and ask back. Same spot, same bubble. Only this time the question belongs to your task alone.

At first it asked the wrong thing

In the first version I made the star a “coach that helps you start.” Look at the content, propose the smallest next physical action you could take right now, in one line.

For a task titled “evening meds,” the star said — “Take out one pill packet and set it on the table.”

Not wrong. A fine nudge for starting the chore. But in that moment I wasn’t taking medicine — I was writing the task. I was filling in the title field, and the star was pushing me to act. The altitude was off.

Not an answer — a question that names the gap

I turned it around. Instead of proposing an action, the star finds one thing in the focused field that’s still empty or vague, and asks you to fill it.

“evening meds” → “Which medicine do you take?” / “When do you take it?”

It doesn’t write the answer for you. It asks one question that makes you surface the next bit of specificity yourself. Where Guide Mode’s static question tells you what the field is about, this one names what’s missing in your version of it.

Short, and sharp

The early questions ran long. The bubble in the top bar is a single line, so long questions got cut off — and English, wrapping word by word, looked worse.

Two fixes. One was the prompt — turning a list of “don’ts” (no casual tone, no parentheses) into “dos,” tightening the length, asking it to be short and sharp. Models, too, follow instructions better than prohibitions. The other was the display — keep the bar’s height, but shrink the text in hint mode so it fits two lines. Long no longer means truncated.

(A small lesson: a reasoning model can burn its whole output budget on reasoning. At first the response came back empty and threw a 502 — but a one-line question needs almost no reasoning, so dropping reasoning to minimal fixed it.)

The price of a little light

Asking an AI costs something. So I decided to call that something Lumen — a sliver of light you spend when you’re stuck.

It used to be called “hint credits.” I renamed it, hoping it becomes a general unit that can power more than hints later. You top it up with Dana — 10 lumen per dana. One lumen each time you ask the star back.

And it’s logged. Down to which field of which task you spent it on — so you can look back at where the light went.

A question, and a better question

A good tool seems to give you answers, but really it gives you better questions. From “what was this field again?” to “ah, I should write down which medicine” — that one field of distance is what keeps the writing moving.

The star was there from the start. It’s just begun to read what you wrote, and ask back.