Chapter 03 · Section V · 9 min read
3.5 Constraints: telling the model what NOT to do
Positive instructions describe the answer you want. Negative instructions rule out the answers you don't. Both matter, and they fail differently.
Most beginner prompts are all positive: “do this, then this, format it like this.” The best production prompts always include a short list of things the model must not do. Constraints are how you close off the specific failure modes you’ve already seen.
Positive vs negative instructions
Positive: “Return a bulleted list of up to five items.”
Negative: “Do not add preamble like ‘Sure, here’s your list.’ Do not include an introduction. Do not use numbered lists.”
They do different jobs. Positive instructions tell the model what shape the right answer takes. Negative instructions tell the model to avoid specific wrong shapes you’ve seen it produce.
When to add negative constraints
Add them when:
- You’ve seen a specific failure and want it to stop. “Do not translate proper nouns” comes after the model translated “Kathmandu” to “the wood city.”
- The default behaviour is unwanted. Most models default to adding pleasantries; “no preamble” fixes that.
- A category of output is off-limits. “Do not include any URLs, phone numbers, or email addresses in the summary.”
- A style is off-brand. “Do not use exclamation marks. Do not use the word ‘unlock.’ Do not start any sentence with ‘In today’s fast-paced world.’”
The four categories of useful negatives
1. Format negatives.
Do not add preamble.
Do not wrap the JSON in code fences.
Do not use bullet points; return one paragraph.2. Content negatives.
Do not invent statistics.
Do not include any personally identifiable information from the transcript.
Do not translate proper nouns or brand names.3. Tone negatives.
Do not use marketing language.
Do not use exclamation marks.
Do not start sentences with "I" or "You."4. Behavioural negatives.
Do not ask clarifying questions; use best judgement.
Do not refuse; if the request is ambiguous, pick the most likely interpretation and note the assumption.
Do not include disclaimers about being an AI.How constraints fail
Negatives are not always followed. Two failure modes:
The Streisand effect. Telling the model “do not mention pandas” occasionally causes it to mention pandas. This is rare with modern models but real, especially with weak or vague negatives.
Over-triggering. “Do not include personal names” can cause the model to redact things that aren’t names, or to refuse the task entirely because it can’t tell what qualifies. Be specific: “Do not include any full names (first + last) from the transcript. First names alone are fine.”
Conflict. “Be thorough” and “keep it short” fight each other. Read your positive and negative constraints together — if they contradict, the model will pick one and you won’t know which.
The “banned words” list
For style-critical work, a short banned-word list is one of the cheapest prompt upgrades:
Style rules:
- Do not use these words: leverage, seamless, robust, best-in-class,
ecosystem, unlock, elevate.
- Do not use "in today's fast-paced world" or any variant.
- Do not use "the world of X."The list is short and specific, so the model can actually follow it. Every entry is one you’ve seen the model produce and don’t want.
Where negatives don’t belong
- When there’s an obvious positive. “Return prose, not bullets” is fine as a negative, but “Return one paragraph, 60 to 80 words” is stronger — it says what to do.
- When they’d need long explanation. If your negative needs a paragraph of “do not do X because of policy Y,” it’s probably better handled by grounding the model in the actual policy (3.3).
Check your understanding
Quick check
—