Length is controlled by structure, not by asking for a word count

The number is not a constraint the way you think

Ask for five hundred words and you may get four hundred or eight hundred. Ask again with emphasis and the spread narrows a little, without closing. This frustrates people out of proportion to its importance, largely because a word count feels like the most concrete instruction it is possible to give.

It isn’t concrete from the inside. Text is produced a piece at a time with no reliable running total, so a target length behaves like a description of the intended size rather than a rule that stops the output. It influences. It doesn’t enforce.

Which means the practical question isn’t how to make the number stick, but what actually determines length. That turns out to be the amount of material and the shape you asked for, both of which you can control precisely.

Structure sets length more reliably than numbers

Three sections of two paragraphs each produces a predictable size. So does five bullet points of one sentence. So does a single paragraph. Specifying the container gets you within a much tighter range than specifying the total, because a paragraph has a natural size and a word count does not.

This is why the instruction that works is usually shaped rather than numeric: one paragraph per finding, no more than four findings, one sentence each on cause and effect. The length falls out of that, and it falls out the same way every time.

It also means that if the output is consistently long, the fix is often to remove a section rather than to lower the number. Asking for eight hundred words across six headings and being surprised at the length is a specification problem, not a compliance problem.

Short is a different task, not a smaller one

Asking for something very short usually produces something generic, because the fastest way to shorten is to remove specifics and keep the frame. The result reads like an executive summary of a piece that does not exist: nothing false, nothing that could not have been guessed.

The way through is to say what the short version must contain rather than how short it must be. One hundred words, and the number and the date must be in them. Now the constraint bites on content, and the generic scaffolding is what gets squeezed out instead.

Very short writing is genuinely difficult and it is worth doing yourself more often than the long version. Cutting a hundred-word paragraph to forty is a series of judgements about what matters, and those judgements are the piece.

Cut afterwards rather than constraining upfront

A workable pattern for anything that has to hit a size is to let the first pass run long, then reduce in a separate step with the target stated. Trimming existing text is a more tractable job than hitting a target while also deciding what to say.

Do the trimming as an explicit operation with a rule attached: remove the least load-bearing sentence in each paragraph, or cut every clause that restates something already said. A rule produces reductions you can inspect. Just make it shorter produces a rewrite wearing the costume of a cut.

Watch what disappears. Reductions tend to strip qualifications, dates and attributions first, because those are the parts that read as optional. They are frequently the parts that made the piece accurate, so the last read before publication should be a check that nothing load-bearing left with the padding.

Where the count actually matters, measure it yourself

Some limits are real: a field with a character cap, a print column, a subject line, an abstract with a stated maximum. In those cases nothing about the request should be trusted, and the count should be taken from the text after it exists.

That measurement belongs to whatever handles the text afterwards, not to the drafting step. Producing three candidate versions and mechanically discarding the ones over the limit is more reliable than any amount of insistence in the wording, and it costs almost nothing.

The general principle is worth keeping. Where a requirement can be checked outside the tool, check it outside the tool, and stop treating the instruction as the place where compliance is guaranteed.

One caveat about hard limits and languages. Character counts behave differently across scripts and across the systems that enforce them, so a version that fits in one field may be rejected by another that counts differently. If a limit matters, test it in the field that will hold the text rather than in a word processor.