Anything only you know is the part the draft will be missing

The gap is predictable, which makes it manageable

Read enough generated drafts on subjects you know well and the same absence appears every time. The general shape is right, the standard considerations are present, the structure is defensible, and there is nothing in it that could only have come from someone who had done the thing. The particular is missing.

This follows directly from where the material comes from. Widely repeated knowledge is well represented; what happened in your own project last spring is not represented at all. There is no wording that recovers information that was never available, and no amount of asking for specificity produces a specific fact rather than a specific-sounding one.

Once you accept that, the tool becomes much easier to use. You stop asking for the part it cannot do, and start using it for the scaffolding while reserving the parts that matter for yourself.

Specific-sounding is the failure mode to watch

Asked for more concrete detail, a draft will happily produce it. Figures appear. A plausible example arrives, complete with a company that sounds real and a percentage that sounds measured. None of it is sourced from anywhere, and the more confidently it is written the more likely it is to survive an inattentive edit.

This is worth calling out because it is the exact opposite of helpful. A vague draft is honestly vague and you know to fill it in. A draft with invented specifics has disguised the gap, and finding those insertions later means checking every concrete claim in the document, which is slower than writing the details yourself would have been.

The practical defence is to never ask for detail, only for the place where detail is needed. Ask for a version with a marker at every point where an example or figure would strengthen the piece. Then fill the markers.

Your material is not always obvious to you

People routinely underestimate what they know that others do not. The specific way your industry handles a common problem, the thing that always goes wrong at a particular stage, the reason a standard piece of advice does not apply in your context — this is the material readers actually want and it feels unremarkable from the inside.

A reasonable way to surface it is to use the tool as a prompt rather than a source. Ask what a piece on the subject would conventionally say, read the list, and mark every point where your experience diverges. The divergences are the article. The rest was already available everywhere.

This inverts the usual workflow in a way that turns out to be productive. Instead of asking for a draft to improve, you are asking for the conventional wisdom in order to have something to argue with, which is a genuinely useful thing to have quickly.

Interviews, notes and the awkward middle

Where the specific material exists but is unstructured — notes from a call, a long thread of decisions, a transcript — the tool is on much stronger ground, because you have supplied what it lacked. Organising, extracting and summarising material you provide is the single most reliable use in writing work.

The care needed here is narrower and easier to apply. Check that a summary has not smoothed away a caveat, that a quoted line is genuinely present in the source, and that nothing has been merged from general knowledge into what was supposed to be a record of what was said. That last one happens quietly and is worth a specific look.

Asking for output that cites its location in the source — a line, a timestamp, a section — makes this check fast enough that you will actually do it. Verification that is expensive gets skipped, and a verification step that gets skipped is not a step.

What this implies about where to spend your time

If the distinguishing material has to come from you, then the reporting, the noticing and the remembering are the parts to protect. The drafting is the part that can be shared out. Most people have the priority backwards, because drafting is the part that felt like work.

It also gives you a straightforward quality test for anything before it goes out. Read it and ask which sentence could not have been written by someone who had never encountered the subject except through other people’s writing. If the honest answer is none of them, the piece has no reason to exist, however well it reads.

That test is uncomfortable and it is the useful one. Fluency was never the scarce resource, and it is considerably less scarce now.