Positive Prompting

Positive prompting is the practice of phrasing prompt instructions in terms of what the model should do (affirmative, constructive directives) rather than relying primarily on lists of what it should not do (prohibitions and negative constraints), based on the well-documented observation that language models — like people following instructions — generally follow clear, concrete positive directives more reliably than they avoid vaguely specified negative ones. A prompt built entirely from prohibitions ("Don't be too formal. Don't use jargon. Don't give a long answer. Don't ignore the user's actual question.") tells the model an ever-growing list of things to avoid without ever clearly specifying what success actually looks like, forcing the model to infer the positive target behavior indirectly; a positively framed equivalent ("Respond in a warm, conversational tone using plain language a non-technical user would understand, in 2-3 sentences, directly addressing the user's specific question.") gives the model a concrete target to aim for directly. This isn't a claim that negative instructions never work — explicit prohibitions remain useful and often necessary, especially for hard safety/compliance boundaries ("Never provide specific medical dosage recommendations") — but production prompt engineering treats "tell the model what to do" as the primary, default framing strategy, reserving explicit "don't" instructions for genuinely hard constraints and boundary cases rather than as the main mechanism for shaping everyday response quality and style. This principle connects directly to output formatting (showing the desired format concretely, rather than just saying "don't be unstructured") and to negative prompts in image generation, where — notably — the positive/negative distinction works somewhat differently, since dedicated negative-prompt fields in diffusion models are a specifically engineered, reliably effective control channel, unlike a wall of "don't" instructions in a text prompt, which are processed as ordinary natural language with the reliability trade-offs described above. Concrete worked example: an AI writing assistant's early prompt read "Don't write like a robot. Don't use corporate jargon. Don't make it too long. Don't be boring." — user feedback showed inconsistent results, since the model had to infer what "not boring" and "not robotic" positively meant, with wide variance across responses. The team rewrote it positively: "Write in a warm, direct, conversational tone, as if explaining to a smart friend. Use short sentences and concrete examples. Keep the response under 150 words. Open with the most important point first." Output consistency, measured via their evaluation set's tone-adherence scoring, improved noticeably, because the model now had explicit positive targets to aim for rather than an open-ended set of things to avoid.