Skip to content

Latest commit

 

History

History
84 lines (58 loc) · 3.33 KB

File metadata and controls

84 lines (58 loc) · 3.33 KB

AI Prompting Guidelines

AI prompting is programming with words, not asking questions. Treat prompts as instructional specifications rather than conversational queries to a probabilistic completion engine.

Core Principles

Use Personas

Assign the AI a specific role and audience context to narrow focus and leverage expertise from a defined perspective.

  • Define who the AI should be (e.g., "You're a senior site reliability engineer")
  • Specify the context of the task
  • This helps the AI draw from relevant domain knowledge

Always Be Contexting (ABC)

Never assume the AI knows something. Missing context gets filled with guesses and hallucinations.

  • Include all relevant facts, even ones that seem obvious
  • Be specific and detailed without holding back information
  • Give explicit permission to say "I don't know" when uncertain
  • The more context provided, the better the output

Define Output Requirements

Specify exactly what you want the output to look like.

  • Format: lists, paragraphs, word count, structure
  • Tone: professional, casual, technical, transparent
  • Style constraints and structural preferences
  • Any specific conventions to follow

Use Few-Shot Examples

Provide 2-3 concrete examples of the desired output style.

  • Teaches through demonstration rather than description
  • Removes ambiguity in expectations
  • Shows the AI exactly what success looks like

Advanced Techniques

Chain of Thought (COT)

Request step-by-step reasoning before arriving at a final answer. This helps the AI work through complex problems methodically.

  • Ask the AI to "think through this step by step"
  • Modern AI platforms often have "Extended Thinking" modes for this purpose

Tree of Thought (TOT)

Request multiple solution approaches to be explored simultaneously.

  • Ask for synthesis and evaluation of the best path
  • Useful for complex problems with multiple valid approaches

Adversarial Validation

Create competing perspectives that critique each other.

  • AI excels at critiquing over original creation
  • Use this to validate and improve initial outputs
  • Have the AI review its own work from a different angle

Process Workflow

Before writing a prompt:

  1. Think clearly about what you actually need
  2. Write specifications in a notebook first if needed
  3. Ask yourself: "Could a human succeed with this information?"
  4. If the answer is no, add more clarity and context

Key Insight

All prompting problems are thinking problems. If your prompt doesn't work, the issue is usually unclear thinking about what you actually want. Clarify your own thinking first; the AI mirrors your clarity level.

References

Academic Papers

Resources

Patrick posts