Describe the shape of the answer before you describe the task

Shape decides content more than wording does

Ask for an assessment of a supplier and you will get prose. Ask for the same assessment as five headed sections — cost, delivery risk, contractual exposure, references, and what remains unknown — and you have not merely reformatted the answer, you have changed what it contains. The fifth heading in particular forces something the prose version would have quietly omitted.

This is the most reliable lever in ordinary use, and it is not about presentation. A structure is a list of questions you have committed to answering. Whatever the shape does not have a place for tends not to appear, and whatever it does have a place for gets addressed even when the material is thin.

That last part cuts both ways, and is worth holding onto for later. A slot that must be filled will get filled.

A named structure beats an adjective every time

Words like concise, thorough, punchy and professional are read against an average that is not yours. Concise might mean four hundred words or eighty. The fix is not a better adjective but a countable instruction: four bullets, at most fifteen words each. Two paragraphs, no headings. One sentence per option, then a recommendation.

Countable constraints are also the only ones you can check without judgement. You can see at a glance whether there are four bullets. Deciding whether something is sufficiently professional takes a reading, and you will make that call differently on a Friday than you did on a Tuesday.

Where a downstream step consumes the output, this stops being a preference and becomes a requirement. Anything a script has to parse needs a shape stated exactly, including what appears when a field has no value, because that is the case that breaks the parsing.

Structure the output so that checking is cheap

The most useful trick in this area has nothing to do with tidiness. If you ask for claims and their supporting evidence in the same paragraph, checking means reading the whole thing carefully. If you ask for them in separate columns, or as a claim line followed by a quoted source line, checking becomes a mechanical comparison you can do quickly and delegate.

The same reasoning supports asking for a short list of assumptions, or of anything the source did not cover, as its own section. It is a place for the tool to put uncertainty rather than smoothing it into the prose, and an empty section on a genuinely thin source is informative in itself.

None of this makes the output more accurate. It makes the inaccuracy easier to find, which in day-to-day work matters considerably more, because the constraint on using any of this at volume is how long verification takes.

Format constrains thinking, for better and for worse

A table forces comparability. That is exactly what you want when the options really are comparable, and it is a problem when they are not, because every cell must contain something and the shape will manufacture an equivalence that does not exist. Options with genuinely different structures get flattened into the same row of criteria.

Bullet points do something similar to argument. An argument depends on connective tissue — because, however, only if, which means that — and a bulleted list strips all of it out, leaving assertions in a vertical stack that look more considered than they are. For anything with reasoning in it, prose is not decoration.

So the choice of shape is a real editorial decision rather than a formatting preference. Choose the one that matches how the material is actually organised, and be suspicious when a neat structure comes back for something you know to be messy.

When to leave the shape open

Early exploration is the obvious exception. If you do not yet know what the answer looks like, imposing a structure will get you a well-organised version of your existing assumptions, and you will have used the tool to confirm rather than to look. Leaving it loose at that stage is a deliberate choice, not laziness.

The transition point comes when you start editing every output the same way. That is the moment the shape has become known and can be stated, and stating it converts a repeated manual step into part of the instruction.

A useful habit is to write the headings before the request. If you cannot name the sections you expect, you have not decided what you are asking for yet, and the output will reflect that with impressive fluency.

There is a middle setting that gets underused, which is to ask for the shape and the answer as two separate requests. Propose a structure first, look at it, adjust the parts that are wrong, and only then ask for the content to be written into it. Two cheap steps replace one expensive round of editing, and a structure is much faster to judge than a finished draft.