The blank page and the last paragraph are not equally good places to ask
Writing has stages with different bottlenecks
Getting started is a problem of not knowing what the piece is. The middle is a problem of thinking — deciding what you believe, in what order, and which parts to leave out. The end is a problem of finish: repetition, clumsy transitions, a sentence that does not parse, a heading that no longer matches the section.
These are not degrees of the same difficulty. They fail differently and they call for different help. Treating a writing task as one undifferentiated job is why people get inconsistent value from the same tool on the same kind of document.
Roughly: strong at the start, weak in the middle, strong at the end. The middle is where the material that only you have is decided, and that is precisely the part that cannot be supplied.
At the start, use it to reduce optionality
The useful early request is rarely for a draft. It is for a list of the angles a piece could take, or the questions a reader would arrive with, or three possible structures with a sentence on what each one prioritises. That converts an unbounded problem into a choice between a few options, which is a much easier problem.
Asking for the conventional treatment is another good early move, not because you want it but because you want to know what you are working against. Seeing the obvious version quickly is worth a great deal when your actual contribution is a departure from it.
The failure to avoid at this stage is accepting the first structure offered. It is generally the most conventional one, since that is what the middle of the distribution provides, and adopting it commits the whole piece before you have thought.
In the middle, it mostly gets in the way
Once the piece is under way, requests for help tend to interrupt the only part of the process that produces anything original. The hesitation over how to phrase a claim is usually a hesitation about whether the claim is true, and handing that hesitation over resolves it in the wrong direction — the phrasing arrives, the doubt is smoothed, and the unexamined claim stays in.
There are narrow exceptions that do not break the flow. Looking up a term, checking whether a sentence is ambiguous, producing a rough version of a passage you already know the content of. All of them are bounded and none of them decide anything.
The broader exception is being genuinely stuck rather than merely slow. When you have stared at a transition for ten minutes, asking for five bad versions is a reasonable way to find the shape of the good one, and you should expect to write it yourself afterwards.
At the end, it is at its most reliable
Finishing work is bounded, checkable and tedious, which is the profile of a task worth handing over. Find the repetitions. List every claim in the piece so you can confirm each one. Point out where a reader loses the thread. Check that the headings still describe what is under them after three rounds of revision.
These requests share a shape: they return observations about text you wrote, and you can verify each observation by looking. That is what makes them safe. Contrast a request to improve the ending, which returns a replacement whose quality you must judge against nothing in particular.
The single most valuable late-stage request is probably the least glamorous. Ask what a hostile or inattentive reader would misunderstand. It surfaces real ambiguity, costs nothing, and catches the sentence you have read fifteen times without noticing that it has two meanings.
Two smaller late requests are worth building into the habit as well. Ask for a plain list of every claim the piece makes, which turns a vague unease about accuracy into something you can work through. And ask which paragraph could be cut with the least loss, because the answer is frequently the one you were most attached to.
Knowing the stage is most of the skill
Almost all of the disappointment reported with these tools in writing work comes from applying an end-stage request at the start, or a start-stage request at the end. Asking for polish on something with no substance yet produces smooth emptiness. Asking for angles when the piece is nearly finished produces a rewrite you did not want.
A workable habit is to name the stage before you write the request, even just to yourself. Am I trying to find out what this is, work out what I think, or clean up what I have? The answer determines what a useful request even looks like.
And the middle stays yours. That is not a limitation to be engineered around; it is where the writing happens, and a workflow that removes it has removed the reason anyone would read the result.
Common questions
Is an outline a good thing to ask for?
As a set of options, yes; as a decision, be careful. Generated outlines tend towards the standard treatment of a subject, which is useful to see and dangerous to adopt unexamined. Take it as a map of what a reader will already expect, then decide deliberately which parts you are keeping.
What about using it to get past a block?
It works, in the specific sense that any external input works: it gives you something to react against. Deliberately asking for a version you will not use is a reasonable trick, because rejecting a concrete attempt is much easier than generating from nothing. Expect to rewrite whatever comes back.
Does this apply to non-prose writing like reports or documentation?
The stages are the same, though the middle is often shorter because the content is given rather than discovered. For documentation in particular the end-stage uses — consistency, completeness against a list, checking that every step is actually stated — are where most of the value sits.
Editor, Prompt After Prompt
Bhavna covers prompt craft, writing with ai, images & audio and the questions readers actually send in and thinks most subjects are more interesting once you know how they work.