1. Enhanced AI Accuracy with Thoughtful Prompts: Research by DeepMind demonstrated that instructing AI to approach problems methodically almost tripled its response accuracy, illustrating the impact of well-crafted prompts.
2. OPRO Technique: This technique involves using AI to instruct another AI, leading to more precise responses and redefining traditional views on AI's input-output quality.
3. Input/Output Model in Prompt Engineering: The article advocates for a structured approach to prompt engineering, focusing on clarity in the role, framework, data, task, and desired output format. This approach helps in obtaining more accurate and relevant results from AI, as demonstrated by various practical examples.
The article highlights the importance of prompt engineering in enhancing AI performance. It showcases how instructing AI to "take a deep breath and think step-by-step" significantly improves accuracy, challenging the "garbage in, garbage out" principle. The technique, termed Optimization by PROmpting (OPRO), involves using one AI to guide another, leading to more accurate responses. The article also explores the input/output model of prompt engineering, providing practical advice on crafting effective prompts to optimize AI outputs, with examples from diverse applications like corporate culture analysis and project management.
AI is a wild frontier.
As you first explore the realm, it feels part science project, part Oregon trail, part Twilight Zone. If you're skeptical about its unpredictable nature, consider this revelation from DeepMind, an Alphabet/Google subsidiary. Researchers found that instructing an AI like ChatGPT to “take a deep breath and work through the problem step-by-step” nearly tripled the accuracy of its response.
OMG. Take a breath and think about that.
This finding questions the age-old principle of "garbage in, garbage out." Through a technique known as Optimization by PROmpting (OPRO), researchers are adopting a meta-strategy to extract more accurate responses from AI. They're essentially having one AI guide another.
When those same researchers subjected an AI to a standard set of high-school math problems, it barely passed, with an accuracy of 34%.
A solid “F.”
When prompted to “think step by step,” the AI got better. The accuracy score leapt to 71.8%. The game changer? Asking the AI to “take a deep breath and think step-by-step.”
Wild.
It’s time to adapt or... get better at prompting.
Generative AI is so startling because it mimics a very human quirk — the ability to learn and adapt. And it does so relentlessly, which should serve as a kick in the pants to all of us. But, like any machine, you get better results if you know how to use it. So it is with AI. The challenge for us isn’t just about getting answers; it’s about asking the right questions.
AI is not just a tool; it's a force multiplier for those who know how to wield it.
Better prompt engineering.
Simple prompts are useful. Type in “TLDR:” before cutting and pasting an email from your boss, and a rambling email turns into a clear explanation.
For a more complete list of handy prompts, here’s a free resource we made.
Or, you could take a course in prompt engineering. DAIR.AI has a great—and free—Prompt Engineering Guide.
The input/output model.
To get the best out of generative AI, use the input/output model of prompt engineering.
At fassforward, we used this model to do something that would have been humanly possible but tedious and terrifically expensive—using AI to accurately describe an organization’s culture.
The AI: In this case, Anthropic’s Claude.
The input: We gave the AI input on how to act (as a business anthropologist) and how to frame the problem (fassforward’s proprietary cultural framework) as well as reams of data to sift through (ten thousand rows of clean, but unstructured verbatim employee feedback).
The output: We specified the output format (a number of positive, neutral, and negative cultural aspects), told Claude what we were NOT interested in, and gave it an example of what an output might look like.
The result: Work that would have taken months, ready in hours. A practical and actionable description of a corporate culture.
While perhaps not as groundbreaking as predicting 3D protein structures, it is a testament to the transformative power of generative AI.
You, too, can write better prompts.
Follow the input/output model. First, the input.
Ask the AI to act as an expert. Are you seeking advice from an expert chef or an award-winning screenwriter?
Next, offer a framework. If you have a specific way of understanding culture or strategy, include it here.
Then, share any crucial data you have. This could be raw data or even a draft you've written.
Task the AI, outlining the work you want it to perform.
And finally, tell it what not to do, such as words or subjects to steer clear of.
Clearly state the output you are looking for.
Format the output you want. Are you looking for a listicle? You want the three different versions? Or the output organized as a table?
Who is your audience? Giving guidance on who’s consuming your output will guide the AI’s level and tone.
Give an example of what you want. This gives the AI something concrete to work from.
Rank or Rate the output in some way.
And remember, it’s a conversation. An ongoing dialogue can only refine the output.
Here’s more detail on each step.
Why speak to a generalist when you can ask an expert?
| Act (Clarify the AIs role)
ChatGPT, Claude, and Bard are general LLMs. But in many cases, you might want the AI to act as an expert... data analyst, marketing strategist, project manager, communicator/ storyteller, HR advisor, leadership coach, or business anthropologist.
Ask the AI to “act as” an expert. This will guide it to a better outcome.
Example prompt line: Act as an expert [Act as an expert...]
Context is king for AIs, too.
| Framework (Offer background details)
Generative AIs, like humans, crave context. It's not just useful; it's vital for effective interaction.
Let's say you've tasked the AI to act as a business strategist. Share your company's goals, challenges, and strategic frameworks like the Balanced Scorecard or Playing to Win. This context sharpens the AI's advice, aligning it with your specific needs. The more you share, the better the AI performs.
Example prompt line: Use the [Framework Name] to guide your work.
Data, useful data.
| Data (Share crucial data)
The devil is in the data. If you've tasked the AI to act as a project manager, the specifics you offer—timelines, available resources, and current project status—are crucial. One of the advantages of AI is you don’t need structured data, but avoid a data dump; take the time to provide clean, curated information.
The more focused your data, the more precise and actionable the AI's project plan will be. This is the foundation for a useful result.
Steering the AI.
| Task (Outline the work for the AI)
The task you outline is the AI's roadmap. For instance, if you've asked the AI to act as an expert sales coach, be specific. Are you looking for cold-calling scripts or strategies to upsell existing clients?
A well-defined task gives you a well-defined answer.
Example prompt line: Make your work [Prompt word] — OR — Give me [Prompt word] answers
Set boundaries.
| Not (Tell the AI what not to do)
Be explicit about what you do not want. That is as important as specifying what you do want. For example, if you've asked the AI to act as a project manager, you might want to specify that you're not interested in methodologies that require a long lead time or high costs. This could be as simple as saying, "Avoid Agile or Six Sigma approaches."
With boundaries, you're not only preventing unwanted outcomes but guiding the AI to spend its computational power on generating the most relevant and useful solutions for you.
Example prompt line: Do NOT reply with [Do not/ Avoid]
Now you have the input, it’s time for the output.
What format would you like?
| Format (Specify the output format)
You can save time by specifying the format you are looking for. This could be a list. You might want a table. Maybe you are looking for step-by-step instructions. In what way is it easiest for you to consume the information?
Example prompt line: Output as a [Output Format]
Make the work relevant to your audience.
| Audience (Identify the target audience)
In a flood of information, relevance beats data. Tune your output to grab the attention of your audience. Specify who the audience is. For example, “produce output suitable for an executive audience.”
This will prompt the AI to be concise and focus on key points. Better yet, simplify and clarify the output. Add “make it understandable for a 6th-grade level.”
Example prompt line: Target your work for an [Audience type] audience.
Provide an example.
| Example (Provide an example)
An example serves as a concrete model for the AI. If you're asking the AI to draft a business proposal, for instance, providing a sample proposal can set the tone, style, and structure you're aiming for.
Think of it as giving the AI a "north star" to guide its output. The closer your example is to what you want, the more useful the AI's output will be. It's like showing a stylist a picture of the haircut you want; what you see is (hopefully) what you'll get.
Rank and rate.
| Ranking (Rate accuracy, relevance, etc...)
Organizing output isn't just about answers; it's about value. Whether you're looking for a list of marketing tactics or project milestones, specifying a ranking or rating provides a nuanced understanding
Ask the AI to prioritize its suggestions based on any number of criteria: feasibility, impact, urgency, confidence. This helps you focus on what's important and gives you a sense of the AI's certainty in each recommendation.
You could also request categorical groupings, like "high-impact but resource-intensive" versus "low-hanging fruit." Think of it as a way to add another layer of utility to the AI's output, making it easier for you to take action.
Example prompt line: Categorize the output by [Ranking and rating] in this [Format].
Chat. Have a conversation.
| Conversation (Maintain the dialogue)
AI has been around for years. What made ChatGPT blow up? I suspect it wasn’t the GPT (Generative Pre-trained Transformer) but the Chat. The ability to have conversations. It lowered the bar for accessibility to AI from a specialized few to the rest of us.
Don’t forget the chat part in your prompting.
You might hear the terms “iterative prompting,” “conversational threading,” or “chain-of-thought” prompting.
Iterative prompting is the back-and-forth between you and ChatGPT (or Bard or Claude). It’s the sense of conversation that will provide more nuance and refinement in each successive answer.
Conversational threading isn’t something you do. It’s the AI. Keeping track of what was said earlier so—like a human—it maintains context in the conversation.
Chain of thought prompting is asking for the AI to work through a step-by-step solution. Particularly helpful when it seems confused, overwhelmed, or going off the rails.
Happy trails.
Remember, it’s a wild frontier. The advice contained here may be horribly out of date in a year.
Meanwhile, take a deep breath, follow the duck, and watch the elephant.
Welcome to Newsfilter AI comments!
Please keep conversations courteous and on-topic.
0 Comments
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere. uis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.