Why an AI assignment planner should ask questions first

Paste an assignment brief into most AI tools and you get a plan back immediately. It looks confident. It is often wrong, because the brief did not say when the draft was due, or whether the bibliography counts towards the word limit, or that your tutor wants primary sources only.

The tool did not know any of that. It filled the gap with an average of every assignment it has ever seen, and handed you a plan for somebody else's coursework.

Guessing is the expensive failure

A wrong plan is worse than no plan. You follow it for a week, and the mistake only surfaces when you are too far in to change direction cheaply. By then you have spent the time.

With no plan, you know you are improvising and you stay alert. With a wrong plan, you have outsourced the alertness to something that was never in a position to be right.

So when Flowstate reads a brief and finds something genuinely ambiguous, it asks you rather than deciding for you. Usually two or three questions, and usually the ones your tutor left out of the brief:

• Is there an interim deadline, or only the final one?
• Roughly how long does one of these normally take you?
• Which parts need a library or lab, so they cannot be done at 11pm?

Answering takes under a minute. What comes back is a plan for your assignment rather than a generic one.

What a clarifying question actually looks like

Abstractly, "it asks questions" sounds like friction. Concretely, it is about forty seconds. Here is a brief of the kind that actually gets handed out:

History Assignment 2 — Source-Based Case Study Choose one of the topics discussed in Term 2. Analyse the reliability of at least three sources. 1,500 words. Due Week 12.

Everything a plan needs is missing. "Week 12" is not a date you can put in a calendar. "Topics discussed in Term 2" is a decision, not an instruction, and it is a decision that has to happen before any of the work can start. Three sources could mean an afternoon in the library or twenty minutes online, depending on what your school counts. And nothing here says whether the analysis is meant to be the bulk of the essay or the last section of it.

A tool that guesses fills all four gaps silently and gives you a schedule that starts with "research sources" on a Monday. So the questions instead:

• Which topic are you doing, or do you still need to choose one? If it is still open, choosing becomes a dated task of its own, sitting early — because a week spent not choosing is the most common way this assignment goes late.
• What date is Week 12, Monday? One answer turns every relative deadline in the brief into a real one.
• Do the sources need to be from the library, or is online fine? This decides whether a task can happen at 11pm or has to be placed on a weekday afternoon.

Answer those and the plan stops being generic:

Nothing in that table is clever. It is just correct for your assignment rather than plausible for an average one, and the difference came entirely from three questions you could answer without thinking hard.

The questions also have to be answerable. "What is your preferred research methodology?" is a question no student should have to field about a 1,500-word essay. The useful ones are the ones you already know the answer to — you simply were never asked.

Where it still gets things wrong

An AI planner that asks questions is better than one that guesses. It is not a solved problem, and it would be strange to publish this without saying where the edges are.

It cannot ask about what it did not notice. The clarifying loop catches ambiguity the model recognises as ambiguity. A brief with an unusual assumption buried in it — a marking rubric that weights one section at 60%, a group component mentioned once in passing — can read as complete when it is not.

Photos of worksheets are harder than pasted text. A clean screenshot is fine. A photo of a printed handout at an angle, in bad light, with handwriting in the margin, is a genuinely harder problem, and the honest expectation is that you will sometimes correct what came out.

It does not know how long things take you. It asks, which helps, but the first few estimates are only as good as your own — and most people are optimistic about their own pace. That improves as you record actual sessions against tasks, and it is not accurate on day one.

Some briefs are vague on purpose. An open-ended research project is meant to make you decide what the question is. No planner should resolve that for you, and the useful thing it can do is make "decide the question" a dated task rather than a thing that quietly consumes three weeks.

None of these are reasons not to build it. They are reasons the plan stays editable, and reasons the last judgement is yours — which is the same principle as the section below.

It plans the work; you do the work

Worth being direct about this: Flowstate breaks a brief into steps with real dates. It does not write your essay, and that is deliberate. The thinking and the writing stay yours — what we are removing is the twenty minutes of staring at a brief trying to work out where to start, not the part where you learn something.

That line matters more for school work than for most software. A tool that quietly does the assignment is a tool you cannot use for the thing the assignment exists for.

What is actually available

Flowstate has not launched. It opens by December 2026, so it is worth being precise about what is committed and what is not:

• Free: full manual planning, study sessions, Pomodoro, streaks, two-way Google Calendar sync, up to 3 active projects or subjects, and one AI breakdown a month.
• Premium (S$6.90/month or S$49/year): unlimited projects and subjects, unlimited AI breakdowns, and AI exam revision scheduling.

A Study Group plan — three to six seats at a discount — is planned future work, not something you can buy. Anything beyond the two lists above is unbuilt, and we would rather say so than list it as a feature.

Why this is the part we are building first

The clarifying loop is why the plan matches the assignment. It is also the part that has to be right before anything else gets built on top of it: every later idea assumes the breakdown can be trusted, and none of them are worth building if it cannot.

So that is where the work is going first, and why we are not promising the rest yet.

In the meantime, the planning method behind the app works on paper too — building a timetable that survives the week and triaging a week where everything is due at once are both things you can do tonight without any software at all.