Every high-quality prompt contains three fundamental pillars. Understanding and consistently applying these pillars transforms vague requests into precise, reliable AI outputs.
The Three Pillars
- Instruction: What you want the AI to DO. This is the core task — summarize, classify, generate, translate, analyze, etc.
- Context: The background information the AI needs to perform the task well. This includes relevant data, constraints, domain knowledge, and user requirements.
- Format: How you want the output structured — JSON, bullet points, table, paragraph, code, etc.
Framework: The ICF Model
| Pillar | Question It Answers | Example |
|---|---|---|
| Instruction | What should the AI do? | "Summarize the following article" |
| Context | What does the AI need to know? | "The article is about quantum computing for a non-technical audience" |
| Format | How should the output look? | "Provide a 3-bullet summary, each under 20 words" |
Good vs Bad Prompts
❌ Bad Prompt:
"Tell me about machine learning"
✅ Good Prompt:
"[Instruction] Explain the three main types of machine learning (supervised, unsupervised, reinforcement).
[Context] The audience is business executives with no technical background. Use analogies from everyday business operations.
[Format] Use a table with columns: Type | How It Works | Business Example | When to Use It"
The Specificity Spectrum
| Level | Example | Quality |
|---|---|---|
| Vague | "Write about dogs" | ❌ Unpredictable |
| Moderate | "Write a blog post about dog training" | ⚠️ Decent but generic |
| Specific | "Write a 500-word blog post about positive reinforcement dog training for first-time puppy owners, in a friendly conversational tone, with 3 actionable tips" | ✅ Precise, reliable |
Practical Tips
- Always start with a clear verb (Summarize, Analyze, Generate, Compare, Classify)
- Provide context BEFORE the instruction when dealing with long contexts
- Specify output format explicitly — never assume the AI will guess correctly
- Include negative constraints ("Do NOT include..." or "Avoid...")
- Test with edge cases to ensure your structure handles exceptions
Common Mistakes
- Writing prompts as if talking to a human ("Hey, could you maybe...") — be direct
- Omitting format specification and getting inconsistent structures
- Providing too much context that obscures the actual instruction
- Not specifying the target audience, leading to wrong complexity level