Why the Word 'Not' Breaks Prompts (Positive Instructions)

1. Concept Overview & Systemic Problem

Every beginner has faced this frustration: you write a prompt to an AI with a slew of strict prohibitions:

• “Write a product description, but DO NOT mention delivery, DO NOT use the word ‘unique’, and DO NOT make the text long.”

And what does the model do in the very first paragraph? It writes: “Our unique product has free delivery...”.

The user gets angry and thinks the bot is mocking them. In reality, this is a fundamental characteristic of autoregressive transformers: to process the word ‘unique’, the model must first load its concept into attention, which automatically increases the likelihood of its output.

For a beginner, understanding this rule is a transition from endless frustrations to precise and predictable control over the outcome.

2. Architectural Taxonomy & Mental Model

┌─────────────────────────────────────────────────────────────┐
│                 HOW TO TRANSFORM A PROHIBITION INTO ACTION  │
├──────────────────────────────┬──────────────────────────────┤
│ ❌ WEAK NEGATIVE PROMPT      │ ✅ STRONG POSITIVE PROMPT     │
├──────────────────────────────┼──────────────────────────────┤
│ “DO NOT write too long”      │ “Write exactly 3 points of    │
│                              │  50 words each”              │
├──────────────────────────────┼──────────────────────────────┤
│ “DO NOT use complex          │ “Write in simple terms for    │
│  scientific terms”           │  a 5th grader”               │
├──────────────────────────────┼──────────────────────────────┤
│ “DO NOT greet and DO NOT     │ “Start the response directly    │
│  write an introduction”      │  with the first line of code”  │
├──────────────────────────────┼──────────────────────────────┤
│ “DO NOT mention our          │ “Talk only about the features  │
│  competitors”                │  of our own product”           │
└──────────────────────────────┴──────────────────────────────┘

3. Technical Pipeline & Internal Mechanics

1. Set Strict Boundaries: Instead of prohibiting excess, clearly outline the limits of what is allowed. For example: “Describe only the technical parameters from the following list: [Your list]”.
2. Use the "Inclusion Filter" Technique: If you need to process a document, write: “Ignore any information except for the last name and phone number”.
3. Provide a Sample of What You WANT to See: Use the Few-Shot technique and show one ideal example without unwanted words.

4. Production Engineering Scenarios

01. E-commerce Product Descriptions

When generating product descriptions, avoid negative prompts. Instead, specify the desired attributes and format to ensure clarity and relevance.

02. Customer Support Chatbots

Instruct chatbots to focus on solutions rather than prohibitions. For instance, instead of saying "Do not mention refunds," guide them to "Emphasize our return policy and customer satisfaction guarantees."

03. Content Generation for Marketing

When creating marketing content, use positive framing to highlight benefits. Instead of "Do not include technical jargon," direct the model to "Use engaging language that appeals to a broad audience."

5. Pitfalls, Common Mistakes & Security

AI is like a small child: if you shout "Do not run through puddles!", they will immediately look at the puddle and run towards it. But if you say "Walk on the dry pavement next to me" — they will calmly walk beside you. Always tell the model WHAT to do, not what to avoid!