The Google Prompt Engineering Whitepaper is excellent, so I created a set of knowledge cards with ChatGPT, π.
π οΈ Best Practices for Effective Prompting
Principle |
Key Idea |
Example / Tip |
Provide Examples |
Use one-shot or few-shot examples to show the model what good output looks like. |
β
Include 3-5 varied examples in classification prompts. |
Design with Simplicity |
Clear, concise, and structured prompts work better than vague or verbose ones. |
β "What should we do in NY?" -> β
"List 3 family attractions in Manhattan." |
Be Specific About Output |
Explicitly define output length, format, tone, or constraints. |
"Write a 3-paragraph summary in JSON format." |
Instructions > Constraints |
Tell the model what to do, not what not to do. |
β
"List top consoles and their makers." vs β "Don't mention video game names." |
Control Token Length |
Use model config or prompt phrasing to limit response length. |
"Explain in 1 sentence" or set token limit. |
Use Variables |
Template prompts for reuse by inserting dynamic values. |
Tell me a fact about {city} |
Experiment with Input Style |
Try different formats: questions, statements, instructions. |
π Compare: "What is X?", "Explain X.", "Write a blog about X." |
Shuffle Classes (Few-Shot) |
Mix up response class order to avoid overfitting to prompt pattern. |
β
Randomize class label order in few-shot tasks. |
Adapt to Model Updates |
LLMs evolve; regularly test and adjust prompts. |
π Re-tune for new Gemini / GPT / Claude versions. |
Experiment with Output Format |
For structured tasks, ask for output in JSON/XML to reduce ambiguity. |
"Return response as valid JSON." |
Document Prompt Iterations |
Keep track of changes and tests for each prompt. |
π Use a table or versioning system. |
π― Core Prompting Techniques
Technique |
Description |
Example Summary |
Zero-Shot |
Ask the model directly without any example. |
π§ "Classify this review as positive/neutral/negative." |
One-Shot |
Provide one example to show expected format/output. |
ποΈ Input + Example -> New input |
Few-Shot |
Provide multiple examples to show a pattern. |
π Use 3-5 varied examples. Helps with parsing, classification, etc. |
System Prompting |
Set high-level task goals and output instructions. |
π οΈ "Return the answer as JSON. Only use uppercase for labels." |
Role Prompting |
Assign a persona or identity to the model. |
π "Act as a travel guide. I'm in Tokyo." |
Contextual Prompting |
Provide relevant background info to guide output. |
π "You're writing for a retro games blog." |
Step-Back Prompting |
Ask a general question first, then solve the specific one. |
π Extract relevant themes -> Use as context -> Ask final question |
Chain of Thought (CoT) |
Ask the model to think step-by-step. Improves reasoning. |
π€ "Let's think step by step." |
Self-Consistency |
Generate multiple CoTs and pick the most common answer. |
π³οΈ Run same CoT prompt multiple times, use majority vote |
Tree of Thoughts (ToT) |
Explore multiple reasoning paths in parallel for more complex problems. |
π³ LLM explores different paths like a decision tree |
ReAct (Reason & Act) |
Mix reasoning + action. Model decides, acts (e.g. via tool/API), observes, and iterates. |
π€ Thought -> Action -> Observation -> Thought |
Automatic Prompting |
Use LLM to generate prompt variants automatically, then evaluate best ones. |
π‘ "Generate 10 ways to say 'Order a small Metallica t-shirt.'" |
βοΈ LLM Output Configuration Essentials
Config Option |
What It Does |
Best Use Cases |
Max Token Length |
Limits response size by number of tokens. |
π¦ Prevent runaway generations, control cost/speed. |
Temperature |
Controls randomness of token selection (0 = deterministic). |
π― 0 for precise answers (e.g., math/code), 0.7+ for creativity. |
Top-K Sampling |
Picks next token from top K probable tokens. |
π¨ Higher K = more diverse output. K=1 = greedy decoding. |
Top-P Sampling |
Picks from smallest set of tokens with cumulative probability β₯ P. |
π‘ Top-P ~0.9-0.95 gives quality + diversity. |
π How These Settings Interact
If You Set... |
Then... |
temperature = 0 |
Top-K/Top-P are ignored. Most probable token is always chosen. |
top-k = 1 |
Like greedy decoding. Temperature/Top-P become irrelevant. |
top-p = 0 |
Only most probable token considered. |
high temperature (e.g. >1) |
Makes Top-K/Top-P dominant. Token sampling becomes more random. |
β
Starting Config Cheat Sheet
Goal |
Temp |
Top-P |
Top-K |
Notes |
π§ Precise Answer |
0 |
Any |
Any |
For logic/math problems, deterministic output |
π οΈ Semi-Creative |
0.2 |
0.95 |
30 |
Balanced, informative output |
π¨ Highly Creative |
0.9 |
0.99 |
40 |
For stories, ideas, writing |
Top comments (0)