Tell the model what to do, not what to avoid

A prompt engineer at an edtech company showed me a system prompt that had grown into a wall of prohibitions. Do not do this. Never mention that. Do not use this tone. Avoid that format. Every time the model did something wrong, the team added another do-not line. The list got longer, and the model kept finding new ways to break the rules, so the list got longer again.

She was frustrated because the effort was clearly not working. More rules were not buying more control. If anything the model seemed to get worse as the list of things to avoid grew.

A prohibition does not describe the target

Here is the problem with a do-not. It tells the model where not to go, and says nothing about where to go instead. “Do not be verbose” rules out one thing and leaves everything else on the table, most of which is also not what you want. The model is left to guess the target from the shape of the hole you cut, and it guesses badly.

There is a subtler cost. Every prohibition names the exact thing you do not want, which puts that thing squarely in front of the model. “Never use technical jargon” is a sentence full of the idea of technical jargon. You have spent words describing the failure in detail and zero words describing success. The model has a vivid picture of what to avoid and a vague one of what to produce.

So a long list of do-nots is a lot of energy spent steering away from things, with no clear picture of the destination. The model keeps drifting, because you told it everything except where you actually wanted it to end up.

Describe the output you want

The rewrite that worked was to turn the negatives into positives. Instead of “do not be verbose,” the prompt said to answer in two or three sentences. Instead of “never use jargon,” it said to write for a first-year student in plain language. Same intent, flipped. Now the model had a target to aim at instead of a fence to avoid.

Positive instructions work better because they are concrete and they point somewhere. A description of the desired output gives the model something to produce. It also tends to be shorter, because one clear statement of the goal replaces the five prohibitions you were using to fence the goal in from every side. When you can, show the shape too. A short example of a good answer does more than a paragraph describing what a bad one looks like.

Ways to steer with positives:

State the goal, not the ban. Say what a good answer looks like, so the model has a target instead of a list of places not to go.
Be concrete about the output. Name the length, the tone, the format you want in plain terms, rather than ruling things out one at a time.
Replace each do-not with a do. When you catch a failure, write the positive instruction that would produce the right thing, instead of adding another prohibition.
Show an example of the good form. Give the model a short sample of the answer you want, since a model imitates a good example more reliably than it obeys a rule.
Keep the few prohibitions that are real. Some hard limits belong as explicit do-nots, so state those clearly and stop trying to fence the whole behavior with negatives.

How we approach it at Density Labs

In the AI Opportunity Assessment, our fixed two-week engagement priced at $2,500 that scopes real production work before you build, we look at how a prompt is written, not just what it is trying to do. A prompt that has turned into a growing list of prohibitions is a common sign that the team is steering by exception instead of describing the target. We rewrite the negatives as positive, concrete instructions and often watch behavior tighten up while the prompt gets shorter. It is a fast change to make early, and a frustrating one to keep patching around once the do-not list has taken over.

A list of things to avoid is a map of the ditches. Give the model the road instead.