Saying what you do not want is the half of the instruction people leave out

A description of what you want excludes very little

Ask for a short explanation of a policy change for a general audience and you have specified perhaps a tenth of the finished thing. The rest arrives from somewhere: an opening line that restates your question, a friendly preamble, three subheadings, bullet points, a closing paragraph that summarises what you just read, and a sentence noting that individual circumstances vary.

None of that was requested and none of it is random. It is the default shape of helpful explanatory writing, which is what the tool has seen most of, and in the absence of any instruction to the contrary it is what you will get. The defaults fill every gap you leave.

This is why the same people write the same three corrections after every single generation. Those corrections are the actual specification, discovered by editing rather than by writing, and they belong in the instruction rather than in your hands afterwards.

Your own edits are the exclusion list

The cheapest way to build a useful set of constraints is to watch what you delete. If you cut the opening restatement every time, that is a line in your prompt. If you always remove the closing summary, another. If you routinely replace three hedged sentences with one direct one, you have found a genuine house rule you had never articulated.

Doing this for a week produces something more valuable than any generic list of prompt tips, because it is calibrated to your output rather than to the average person’s. It is also stable — these defaults do not change much between tools, since they come from the same underlying pool of instructional writing.

Keep it short enough to stay readable. A prompt with fifteen prohibitions is unwieldy and the later ones tend to lose force. Six or seven that address things you genuinely fix by hand will do more than an exhaustive list copied from somewhere.

A prohibition works better rewritten as an instruction

Telling a tool not to be flowery is weak, because it names a quality without supplying a replacement. Telling it to use plain words, keep metaphors out, and prefer a concrete example to an adjective gives it something to do instead. The behaviour you want has to be reachable, not merely the behaviour you want avoided.

The same applies to structure. "No bullet points" often produces prose with the rhythm of bullet points, a series of short declarative sentences that were clearly a list a moment ago. "Continuous prose, at least three sentences per paragraph" gets you where you were going, because it describes the destination.

This is not a quirk of the technology so much as a general property of instructions. A negative constrains a large space to a slightly smaller large space. A positive picks a point in it. Where you can convert one into the other, do.

Where negation genuinely does not hold

Image generation is the clear case. Asking for a street scene without cars frequently produces cars, because the words in the request all pull towards the same region regardless of the grammar attached to them. Most image tools handle exclusions through a separate mechanism precisely because writing them into the description does not reliably work.

Long chains of prohibitions have a second failure mode worth knowing. Every item you mention raises its salience, so a prompt spending four sentences on what to avoid has spent four sentences describing that material in detail. Occasionally the thing you forbade appears more often, not less.

The practical rule is to prohibit sparingly and specifically, and to reach for a positive alternative whenever one exists. Where an exclusion is genuinely important — a term that must never appear, a claim that must not be made — put it in its own short line rather than in a paragraph of general guidance.

The parts worth being strict about

A small number of exclusions are worth stating even when they feel obvious, because the cost of the failure is high and the cost of the sentence is nothing. Do not invent a figure, a name or a citation. Do not state anything the supplied source does not support. If the source is silent, say so instead of filling the gap.

These will not be honoured perfectly, and treating them as a guarantee is the mistake. What they do is shift the tendency, and more usefully, they give you something to check against: you asked for gaps to be flagged, so an output with no flags on a thin source is itself a signal worth a second look.

The rest is taste, and taste is where the specification is really being written. Everyone assembling this list ends up with a slightly different one, which is a sign it is doing its job rather than a sign of inconsistency.