A Guide to Prompting: What You'll Learn
- Prompting best practices for getting what you want from ChatGPT.
- Discover practical examples for everyday tasks.
- Troubleshoot common prompting challenges.
- Pro-tips & resources for continued learning.
I'll share some of my favorite example prompts; use them! But the true goal of this post is for you to understand why they work.
Why Prompting Matters
As Search marketers, we've understandably been trained to use and understand search engines; but AI language models require a different approach. One day natural language alone may be sufficient, but for now, effective prompting can seriously improve the quality of what you get. Remember, these models can only give you what you ask for. If you don’t ask the right questions, you don’t get the right answers. There’s a skill in crafting the right prompts, and that’s what we’re going to focus on today.
There's a lot of advice out there about how to structure your prompts. Do you need to give it a role? Do you really need to become a "prompt engineer"?
I'd like you to walk away understanding why these tips and tricks work, more so than just memorizing a bunch of rules. Because let's face it, over time, this is going to change. The models are going to be able to help us do a lot of this prompting themselves - and I'll show you a couple of tricks along the way to get them to do that for you now.
The AI: From Panic to Progress
When OpenAI's CEO, Sam Altman, said that 95% of what marketers do could be handled by AI, it sped up the emotional rollercoaster that we've been on since ChatGPT launched..
Panic: "Oh no, AI is going to take our jobs!"
Denial: "AI-generated content is terrible. It'll never replace human creativity."
I'm sure you've seen it - bad content generated by AI, overusing words like "synergy" and "delve" - it's beyond average.
The problem with denial, is that there are multiple studies showing incredible gains in productivity from integrating AI into our daily workflows.
A Harvard study showed that marketing professionals leveraging AI tools can demonstrate significant productivity. On average, AI-empowered consultants completed 12% more assignments, reduced task completion time by 25%, and delivered work of 40% higher caliber compared to their counterparts not using AI assistance.
These gains are not something to panic about - that's something to get pumped about.
Acceptance: "Okay, AI actually can really help us. How can we use it effectively?"
AI Needs Subject Matter Experts to Work Well
But here's the thing: studies are showing some impressive gains when marketers effectively use AI. For instance, consultants using AI finished 12% more tasks, completed them 25% more quickly, and produced 40% higher quality results. That's not something to panic about - that's something to get excited about!
Why Prompting Matters More Than You Think
We've all been trained to use search engines, right? But here's the kicker: AI language models require a whole different approach. Many folks think that just typing in natural language is enough, but trust me, there's so much more to it.
Here's the problem: searches make pretty awful prompts. "Why is my website not getting traffic" or "best types of Facebook posts" - these are really not great prompts. They're very generic. Google's gotten pretty good at feeding you information with these, but LLMs? Not so much.
And here's something crucial to remember: LLMs are not search engines. I love this slide from Britney Muller. It's such an important reminder. LLMs are not information retrieval systems. They're not deterministic. It's called generative AI. They generate randomized outputs based on a probability distribution.
So, because of that, AI really does need subject matter experts to work well. There's no instruction manual on some tasks. AI is immensely powerful on others it fails. But unless you use AI a lot, you're not going to know which is which. And this is a skill that you can build.
As Ethan Mollick puts it, prompting is "a little bit of magic, but mostly just practice." That's exactly right - it's a skill you can develop with time and experience.
Working With AI is a Skill You Can Build
Principles of Effective Prompting
Let's break down the key principles that'll help you get the most out of your AI interactions:
1. Be Clear & Specific
When it comes to prompting, clarity is king. Here are some tactics that'll help:
- Use Delimiters: Throw in some """...""", '''...''', or --- to separate your instructions and examples. It helps the AI understand exactly what you're asking for. Break up your text on new lines with shift+enter.
- Ask for Structured Output: Want your response in Markdown or HTML? Just say so!
- Check Assumptions: Make sure the AI verifies its assumptions. It'll help avoid those "Wait, what?" moments when you're reviewing the output.
- Few-shot Prompts: Give the AI some examples of what you're looking for. It's like showing a new employee exactly how you want a task done.
Here's an example prompt that puts these tactics to work:
Summarize the meeting notes in a single paragraph. Then write a markdown list of the speakers and each of their key points. Finally, list the next steps or action items suggested by the speakers, if any. Check your work for completeness.
Here's a good example:
"""
[Paste in an old recap]
2. Give Time to Think
We all know rushing leads to mistakes. The same goes for AI. Here's how to give your AI tool the time it needs:
- Chain of Reasoning: Ask for a step-by-step plan. It helps the AI organize its thoughts (and gives you better results).
- Specify Steps: Clearly outline the steps you want the AI to take. It's like giving it a plan to follow.
- Encourage Solution Generation: Let the AI work out its own solution without rushing to conclusions. You might be surprised by what it comes up with!
I'm launching a new product line and need site content. Outline a step-by-step plan for launching a new product in the market. Begin with market research, then product development, then marketing strategy, and end with post-launch analysis.
3. Iterate
Your first prompt probably will almost never give you exactly what you want. That's okay! Here's how to refine your approach:
- Analyze and Refine: Take a look at why the initial response wasn't quite right. Remember the first two principles. Ask for steps and clarify your instructions.
- Provide Examples: Use batches of examples to refine your prompts. It's like training the AI to understand your specific needs.
- Collaborate with AI: Make ChatGPT ask you for the information it needs..
Reply again, this time:
- Write a step-by-step plan before proceeding,
- Bullet list important points
- Use the files I sent as your primary source
- Search online & cite all sources, NO assumptions
- Ask me for necessary follow-up information
Overcoming Common Challenges
ChatGPT Can't Read My Data (CSVs, etc.)
Data wrong or returning errors? LLMs aren't optimized to 'talk to data' like CSVs, databases, etc. But - they are good at coding - and code is good for data work!
Here are ways to work with data in ChatGPT:
- Use Python or Code Interpreter: Specifically ask the LLM to use Python or its code interpreter for data tasks.
- Work in Small Batches: Break down data tasks into manageable parts; this gives LLMs "time to think!"
- Require Completion: Tell it "do not truncate or use placeholder text" to get complete responses.
- Provide Examples: Offer example output files like CSV templates.
- Try turning it off and on again: Try a new chat - pro-tip, paste the old one in as Markdown!
Example Prompt for Website Categorization
- Start by categorizing the first 20 websites in my list, then prompt me for feedback before the next batch
- Always fully implement all requested functionality. NO placeholders or todos. All code MUST be fully written and implemented.
LLM's still struggle with web crawling, huge datasets, and generating full slide decks. For those tasks:
- Web crawling
- use Screaming Frog
- Cleaning large datasets
- transform/pre-clean if possible
- Slidedeck generation
- Use LLM for outlines rather than slides
ChatGPT Keeps Hallucinating
Another common issue that leads to "denial" are hallucinations - when AI just makes fake things up that sound true.
So how do we keep the AI grounded?
Well, first we have to set our expectations properly; recall that LLMs aren't search engines and will never retrieve exactly correct information.
Remember - generation is the feature, not the bug! If you flip perspective, you could say that a search engine has a creativity problem; it never makes anything new. This tweet from Andrej Karpathy explains it well:
That said, there are a few ways to increase accuracy.
- Use Python for Math: Make sure mathematical tasks are solved step-by-step using Python to avoid errors; coding is more accurate than generation.
- Use R.A.G.: Utilize Retrieval Augmented Generation to constrain responses and improve accuracy.
R.A.G. is an especially critical concept. An easy way to break it down is into three parts:
- Retrieve information from a trusted reference text
- Augment replies with citations from references
- Generate replies constrained to those references
What this means, is you're telling ChatGPT to pull in reference texts and only give information from within those references text. So, basically, saying limit your generations!
This approach helps ground the AI's responses in factual, up-to-date information, reducing hallucinations and improving accuracy.
I'm Not Getting the Responses I Want
Don't worry, we can fix that:
- Ask for Variations: Request multiple variations and select the best. LLMs never get tired, so keep asking!
- Force Iteration: Make the AI prompt you for further questions to refine responses. This is my absolute favorite prompt (from Seth Waite):
Ask me any questions you need to know to understand the problem more. After each set of questions, tell me on a scale of 0 to 100 how confident you are in the answer. Don't try to tell me the answer until you're at least 95% confident.
One important tip: don't ask the AI if something is correct. It will try to predict what you want rather than give you an accurate answer. Instead, ask it to work through the problem step-by-step and show its reasoning.
Advanced Tips & Resources
Custom GPTs: Let AI Do the Prompting
There are lots of custom GPTs that are good at developing prompts. Here are some of my favorites:
- Prompt Redesigner: I use this all the time. I put in a prompt, call in the prompt redesigner, and it does a better job for me.
- AutoExpert (Chat): This tries to figure out what kind of expert would be best to handle your problem.
- GPT Instruction Builder: Helps build out custom GPT instructions.
- Grimoire: A really nice coding GPT that remembers to pull in necessary Python libraries.
- No More Delve: Great for auditing content and removing overused phrases.
- Super Describe, Image Edit / img2img, and Cauldron: These are fantastic for working with images and generating prompts based on visuals.
Try Other Models
If you're in ChatGPT and your prompt isn't quite giving you what you want, there are other free models out there:
- Gemini 1.5 2M: Available through Google's AI studio.
- Claude 3.5: Anthropic's best model, available for free.
- Llama-3: Meta's top model, also freely accessible.
And if you want to play around with multiple models, Poe.com offers access to various models for $20/month.
I Love ChatGPT Memory & Custom Instructions
I absolutely love ChatGPT's memory and custom instructions functions because you can improve your default responses. I put some of my favorite prompts in there, like having it ask me for questions when I prompt it every time, or telling it to always avoid using certain words.
Useful Tools for Prompt Engineering
When working with prompts and AI-generated content, these tools can be super helpful:
- MarkDownload Chrome Extension: Easily copy web pages as Markdown.
- Paste into Markdown: Convert copied text into Markdown format.
- Markdown to HTML / Rich Text Converter: Transform your Markdown into other formats as needed.
Additional Resources
To help you on your AI prompting journey, here are some valuable resources:
Prompt Libraries
- Prompt Library — More Useful Things: AI Resources
- Prompt Library - Anthropic Claude
- GPTs Prompts Leaked List
Prompting Guides
- Prompt Engineering overview - Anthropic
- Prompt Engineering - OpenAI
- Key Guidelines for Writing Instructions for Custom GPTs | OpenAI Help Center
- Gemini Prompting 101 Guide
Prompting Courses
Wrapping It Up:
Remember, mastering effective prompting techniques is an ongoing process. Keep experimenting, keep learning, and don't be afraid to push the boundaries of what these AI tools can do.