Negative Prompting: Why "Don't Do This" Beats Detailed Descriptions
That is, try a simple test. Ask a language model to write a chapter about “loneliness” that is “intimate, deep, literary, fluent, and moving.” The result, in most cases, is a text that has all five qualities at once and none of them: tidy, safe, and utterly forgettable sentences. Now repeat the same request with three prohibitions: “Do not use the word loneliness more than three times. Do not begin any paragraph with ‘in fact.’ Do not write a moral conclusion at the end of the chapter.” The gap between these two outputs is often vast, and the reason is neither luck nor the length of the prompt.
Adjectives Make Wishes; Prohibitions Make Conditions
A language model cannot assess whether its text has become “deep.” Such words have no criterion for it, so at best it moves toward the average of its trained style. And the average is exactly where clichés live. When you write “make the prose more literary,” the model moves toward the most conventional definition of “literary”: ready-made metaphor, repeated similes, a rhythm we have heard a thousand times.
By contrast, “do not use this word” is a binary operation. The model can check whether it has complied, and so can you. A positive instruction is a wish; a negative instruction is a verifiable condition. When you put ten conditions of this kind together, the model falls off the default highway and is forced to seek a path it has taken less often in its training data—and that path is usually fresher.
Oulipo: When Prohibition Became an Engine of Creation
This idea has a long history in literature. Georges Perec wrote roughly three hundred pages in the novel La Disparition (1969) without using the letter “e” even once; a book that Gilbert Adair later translated into English under the same constraint. The letter ban deprived the writer of all the easy paths and led him to vocabulary he himself did not know. Raymond Queneau, too, in Exercises in Style, narrated an insignificant incident dozens of times under different constraints to show that constraint is not the enemy but the source of variety.
Great twentieth-century writers also wrote their lists mostly with “don’ts.” Elmore Leonard, in his famous ten rules, says do not start a book with the weather and delete the adverbs; George Orwell, in “Politics and the English Language,” follows almost the same path and warns against the worn-out metaphor and the big word in place of the small one. Even the famous rule “delete the unnecessary words” is a fully fledged negative instruction.
What Should We Prohibit in a Book Chapter?
The following list comes from the experience of generating chapters with language models; each row is a recurring cliché that is curbed by a short instruction.
| Prohibition | Why It Works |
|---|---|
| Starting the chapter with a definition of the topic (“In today’s world…”) | The adult reader has seen this sentence hundreds of times and rejected it right there |
| Emotional adverbs (“she said sadly”) | The scene should create the feeling itself; the same lesson as Hemingway’s iceberg theory |
| A moral summary in the final paragraph | The reader should construct the meaning, not have the writer deliver it |
| Lists of three and consecutive parallel sentences | The rhythm becomes mechanical and predictable |
| Repeating the chapter’s key word more than three times | The key word fills the space of thought with itself |
| Sentences longer than three lines | It leaves the Persian reader breathless and increases ambiguity |
How to Write a Healthy Negative Prompt
- Write the prohibition at the level of structure, not the level of the word. A model follows avoidance of a structure better than avoidance of a word.
- For each prohibition, provide a short alternative; otherwise the model fills the gap with something that is easier for it.
- Keep the number under ten, and put two or three critical items first in the list.
- Tie the constraint to a specific chapter, not to the whole book; a character-driven chapter does not have the same prohibitions as an analysis-driven chapter.
Instead of “write the chapter in a warm and human way,” write: “Do not begin any paragraph with a definition. Do not use emotional adverbs. Do not include a conclusion at the end. Do not create sentences longer than three lines.”
Pitfalls: The Pink Elephant and Endless Prohibitions
A technical warning is also necessary. When you name a word in the prompt, that word enters the context and may increase the likelihood of its appearing; something like “don’t think of a pink elephant.” So for individual words, it is better to use alternatives (“instead of the word loneliness, write the scene of an empty room”) and reserve prohibitions for structures and patterns.
The second pitfall is an accumulation of constraints. A list of twenty prohibitions makes the model rigid and conservative, and the result is text that has no errors and nothing else either. The third pitfall is performative prohibitions; conditions the model would never violate and that only increase the length of the prompt. Before adding any “don’t,” ask yourself once: Have I seen this error in previous outputs? If not, remove it.
Conclusion
Negative prompting does not take the place of positive prompting; it maps the route positively, and prohibitions merely close off side roads and repetitive paths. But in generating a book chapter, it is precisely this closing of paths that pulls the text away from the average. If saying “don’t do these things” seems more effective than precisely describing what you want, it is because negation creates a standard, while wishing does not.
