Immediately improve the output you receive from AI models

1. Use the playground or workbench

• Stop using mass market facing software wrappers
• Learn to use the playground sections of AI models for better control and input
• Direct API access gives you more control over parameters and system prompts

2. Model performance decreases with prompt length

• The longer your prompt, the worse the response
• Do not remove the rules and instructions, just shrink them
  • KISS principle applies
  • Ideal: 250 tokens or at least <500
• Try wordcounter.net to track token count
• Simplify, reread, and simplify again
• Stop adding unnecessary fluff to your prompts

3. System, user, & assistant

Understand the different prompt types

• System level prompts (High level instructions)

  • “You’re a helpful, intelligent assistant”
  • “You’re a cybersecurity expert”
  • “You’re a highly successful CTO”
  • Sets the overall behavior and expertise level

• User level prompts (The actual instruction)

  • Tell the model what you want it to do
  • Be specific about the task and desired outcome

• Assistant level prompts (Building on previous outputs)

  • Use the output from the model, encourage it (if good), then build from it
  • “Layering and reinforcing” - the assistant’s output becomes part of the context window
  • Example: “Fantastic, now I want you to do the same thing, but for [specific scenario]”
  • Threading and reinforcing previous responses

4. Use one-shot or few-shot prompting

• Always use a minimum of 1 example in your prompts
• Show the model exactly what you want with concrete examples
• More examples = better results, but balance with token limits

5. Conversational vs Knowledge engines

• LLMs are typically conversational engines
• Tables and databases are knowledge engines
• LLM table/db/encyclopedia
  • RAG (Retrieval-Augmented Generation) bridges this gap

6. Use completely unambiguous language

• You want the perfect output “zone”
• Bad example: “Produce me a report based on this data”
• Better example: “List our 10 most popular products and write a one paragraph description of each”
• Even better example: “List our 10 most popular products and write a one paragraph description of each. Here is an ideal example of the output I want: [specific example]“

7. Spartan tone

“Use a spartan tone of voice”

• When trying to specify a good tone of voice, use this as a good place to start
• Clear, direct, and to the point without unnecessary elaboration

8. Iterate prompts with data

• Or, throw spaghetti at the wall, repeatedly
• Group similar prompts and analyze patterns
• Reduce variation to find what works consistently

Try with a spreadsheet

1. Prompt
2. Output Do this ten times
3. Then determine what is good enough How many of the 10 are good?

9. Define the output format

• Output JSON
• Output Markdown
• Output bullet points
• Code blocks
• Specify the exact structure you want

10. Remove conflicting instructions

• “Detailed summary” is conflicting - is it detailed or is it a summary?
• Do you want it user-friendly or do you want it accurate?
• Do you want it engaging or do you want it comprehensive?
• Choose one direction and stick to it

11. Learn to use JSON, XML, and CSV

Ways to structure data

• XML = Extensible Markup Language

<author>ZeroDumb</author>
<date>2025-08-04</date>

• JSON = JavaScript Object Notation

{
	"author": "ZeroDumb", 
	"date": "2025-08-04",
	"hobbies": ["cyber security", "hackery", "reading"], 
	"location": "your root domain",
	"aliases": [ 
		{"name": "mmaddhatter13"},
		{"name": "Zero"},
		{"name": "ZeroDumb"}
	] 
}

• CSV = Comma Separated Values

author,date,title
ZeroDumb,2025-08-04,prompting

Note: LLMs may lose their place in large CSV sheets

12. Key Prompt Structure

• Context: What the model needs to know
• Instructions: What you want it to do
• Output format: How you want it structured
• Rules: Constraints and limitations
• Example: Concrete demonstration

Example: “You’re an intelligent admin that filters jobs. Given the following job posting, determine if it’s suitable for a junior developer…“

13. Use AI to create examples

• You don’t have to have 8 examples, you can use AI to help generate examples
• Start with one good example, then ask the AI to create variations

14. Use the correct model

• Always try to use the best model available for the job you are asking it to do
• Consider model capabilities, context limits, and cost when choosing

15. Test and iterate

• Prompt engineering is iterative
• Test your prompts with different inputs
• Measure success rates and refine accordingly
• Keep a prompt library of what works

---

Question loudly so others can learn quietly. Stay curious. Stay loud.

Don’t Be A Skid -Zero

Buy Me A Coffee @iamnotaskid