7 Ways To Stand Out With Your ChatGPT Prompts

Key Concepts

• Prompt Engineering: Crafting clear and effective prompts for Large Language Models (LLMs) like ChatGPT.
• Lost in the Middle Phenomenon: LLMs tend to lose context in the middle of long conversations.
• Context Management: Strategies for providing LLMs with relevant information and instructions.
• Instruction Levels: System instructions, custom instructions, and chat-level instructions.
• Feature Utilization: Leveraging specific features of LLMs like Canvas, Deep Research, and Agent Mode.
• Header-Based Communication: Structuring prompts with clear headers and formatting.
• Explore vs. Execute: Distinguishing between brainstorming and task execution phases.
• Generate vs. Critique: Using LLMs for both content creation and content improvement.
• Few-Shot Learning: Providing LLMs with examples to guide their output.

1. Edit, Don't Pile On: The U-Shaped Context

• Main Point: Instead of continuously adding clarifications in a long chat, edit the original instruction.
• Lost in the Middle Phenomenon: Research from the University of Washington, MIT, and Google Cloud AI research shows LLMs have a "U-shaped" context window, remembering the beginning and end of a conversation better than the middle.
• Example: Instead of saying "Summarize the book 'Notes on Complexity'," then "Make it shorter," then "Use bullet points," edit the original instruction to say "Summarize the book 'Notes on Complexity' briefly, no more than three paragraphs, and use bullets."
• Actionable Insight: Regularly revisit and refine your initial prompt to maintain context and improve output quality.

2. Projects and Spaces: Contextualizing Conversations

• Main Point: Use projects (ChatGPT paid feature) or spaces (Perplexity free feature) to provide persistent context to the LLM.
• Projects (ChatGPT): Allow uploading files (e.g., books, documents, transcripts) and setting custom instructions that apply to all chats within the project.
• Example: Upload the book "E-Myth Revisited," a YouTube video script, and a business coaching discovery document to a project. Then, ask ChatGPT to suggest workflows based on the uploaded context.
• Custom Instructions (Projects): Define how ChatGPT should interact with you within the project (e.g., organize ideas, analyze priorities, challenge inconsistencies).
• Actionable Insight: Organize your knowledge and notes effectively to provide LLMs with relevant context for more personalized and accurate responses.

3. Instruction Levels: System, Custom, and Chat

• Main Point: Understand and utilize the different levels of instructions that influence LLM behavior.
• System Instructions: Set by OpenAI, defining the overall behavior of ChatGPT (e.g., "helpful assistant").
• Custom Instructions (Profile Level): Set in your ChatGPT settings to define your preferences (e.g., preferred tone, format, level of detail).
• Custom Instructions (Project Level): Set within a project to define how ChatGPT should interact with you specifically for that project (e.g., act as a coach, organize ideas).
• Example: In custom instructions, specify that you want ChatGPT to provide information with cited studies, URLs, anecdotes, and cross-domain thinking.
• Actionable Insight: Customize your profile and project instructions to tailor ChatGPT's behavior to your specific needs and preferences.

4. Ask for Features: Canvas, Deep Research, Agent Mode

• Main Point: Explicitly request specific features of the LLM to achieve desired results.
• Canvas: Creates a collaborative document where you can directly edit and modify the LLM's output.
• Deep Research/Agent Mode: Enables the LLM to conduct in-depth research and analysis.
• Example: Instead of just asking for a summary, say "Give me a canvas of a summary of the book 'Notes on Complexity'."
• Actionable Insight: Explore and utilize the various features offered by LLMs to enhance your workflow and output quality.

5. Talk with Headers: Structured Communication

• Main Point: Structure your prompts with clear headers and formatting to guide the LLM's response.
• Example: Instead of a vague request, specify "Give me the core idea, counterintuitive ideas, problem + example, and solution + steps" as headers.
• Benefits: Improves clarity, focuses the LLM's attention, and facilitates easier editing and refinement.
• Actionable Insight: Use headers and formatting to organize your prompts and make it easier for the LLM to understand your requirements.

6. Explore vs. Execute: Defining Your Goal

• Main Point: Determine whether you are in the exploration (brainstorming) or execution (task completion) phase and adjust your prompts accordingly.
• Explore Phase: Use vague prompts and ask for multiple options (e.g., "Give me 10 options").
• Execute Phase: Use specific prompts with clear goals, context, and tasks (e.g., "Create a YouTube script with a 5-minute format and business casual tone").
• Generate vs. Critique: Use LLMs not only to generate content but also to critique and improve existing content.
• Example: Ask ChatGPT to critique a YouTube script and provide principles and rules for making it more engaging, citing psychology and behavioral science.
• Actionable Insight: Clearly define your goals and adjust your prompts to match the exploration or execution phase you are in.

7. Examples: One-Shot and Few-Shot Learning

• Main Point: Provide examples to guide the LLM's output, especially when you have difficulty articulating the desired tone or style.
• One-Shot Learning: Provide one example of the desired output.
• Few-Shot Learning: Provide multiple examples, including both positive and negative examples.
• Example: Instead of asking for a "playful poem," provide an example like "The Little Prince was like a stubborn puppy."
• Actionable Insight: Use examples to communicate nuanced requirements and improve the LLM's ability to generate content that matches your specific needs.

Conclusion

The key to effective ChatGPT usage lies in understanding and applying prompt engineering principles. By editing prompts instead of piling on, leveraging projects for context, utilizing instruction levels, requesting specific features, structuring communication with headers, distinguishing between exploration and execution, and providing examples, users can significantly improve the quality and consistency of ChatGPT's output. Ultimately, bad prompting stems from a lack of clarity about what you want, and by addressing this, you can unlock the full potential of LLMs.