The Evolution of GenAI LLM Prompt Engineering: From Zero-Shot to Chain of Thought

Remember when ChatGPT burst onto the scene, and we all jumped on the bandwagon firing off zero-shot prompts like “Write me a blog post on why cloud marketplaces are the future of tech B2B sales”… and then being grumpy about the terrible results? The disappointment was worse than being a Leeds fan us to score from a corner.

But oh, how far we’ve come! (Though, let’s not get over our skis… there’s too much hype out there… let’s keep it real… but there is great value in LLMs).

Take a stroll with me down memory lane. Let’s see how we’ve evolved from those early days of clumsy prompts and disappointing answers. Those feelings are now replaced with “I’m a bit nervous but slightly excited” era of chain-of-thought reasoning embedded in consumer-focused LLMs like ChatGPT o1.

This article is part of the coming-soon Viewyonder project, Finding your GenAI Happy Path. The goal is to help bring everyone up to the same level: not everyone uses LLMs or reads the white papers as much as saddos like me. I know some people have barely scratched the surface of LLMs, or have got suboptimal results — and, if I’m honest, I think you’re missing out and I want you in the party with me.

Following this article is an example of a latest practice you can try yourself. If you improve it, share back via our social or Contact.

• LinkedIn
• X
• Facebook
• Mail

1. The Research Grind

Remember the early days when we grumbled because LLMs weren’t connected to the internet? We still had to lace up our boots, grab a brolly, and walk the dog (as we say in the biz) to do the research ourselves. It felt like we were living in the stone age and our first version of a wheel was square and didn’t work, but we knew something was almost right about it…

💡 Fun Fact: Perplexity.ai emerged as an early attempt to automate the research process, combining the power of LLMs with real-time web search capabilities.

2. From Raw Data to Polished Copy

As we got more comfortable with LLMs, we realized we could feed them our “raw research” and ask them to whip up some tasty content based on our findings. Suddenly, our results started looking a lot more appetizing!

But there were problems. Every article started with the same passive-voice “In the age of…” and used words like “delve”. It’s not quite right, just yet. How can we improve it?

3. The Birth of Prompt Engineering

Enter stage left: Prompt Engineering!

We discovered that giving our LLM buddies better guidelines resulted in content that was better. This was when “prompt engineer” became the hottest job title since “social media guru.”

Instead of zero-shot “write me a thing about cats”, we instead invested more effort in the prompt. It became more like the brief that you would give to a real writer. Use this tone of voice. I want top 5 cat breeds. Leave out the bald ones. That kind of thing. 😜

📊 By the Numbers: itjobswatch.co.uk reports today that “prompt engineering” is a tiny proportion of IT jobs, but it’s doubling every six months. A quick search for “prompt engineer” jobs at the time of writing showed 574 in the UK, and they are across industries from agencies to telecoms to insurance and legal. I expect prompt engineering to become a universal skill, not a job in its own right — that’s kinda the purpose of this article.

4. Show, Don’t Tell: The Power of Examples

Remember that writing advice your high school English teacher drilled into your head? “Show, don’t tell”? Turns out, it works for LLMs too. We learned two killer techniques:

1. Paste an example directly into the prompt.
2. Get the LLM to reverse-engineer a good example and turn its analysis into prompts.

For example, I once had an LLM reverse-engineer a snarky friend’s post into guidelines, then wrote an article using that: and we did, indeed, get a snarky article (but not as good as one my friend’s originals, because it was missing all the experience, data, and “connections” my friend has… there’s a clue to the future here…).

Suddenly, our LLMs were producing content that could make Hemingway weep into his daiquiris.

5. LLMs in Disguise: Role-Playing for Better Results

But what if we go the LLM to assume the “personality” of the guidelines we just got it to reverse engineer?

We took things up a notch by telling our LLM wingman buddy-thingy to pretend they were subject matter experts or copywriting geniuses and gave them the guidelines to be that role. Like a job spec.

We even gave them “their own writing samples” (which were actually penned by the likes of Hemingway or Paul Graham).

🎭 Pro Tip: When asking an LLM to mimic a style, try this prompt: “Analyze the writing style of [Author Name] and explain the key elements that make their writing unique. Then, use those elements to write in their style.”

6. Divide and Conquer: Multi-Step Prompting

But if that person was real, you wouldn’t give them a one-liner like a joke out of a Christmas cracker and expect great things, would you? If you’ve ever written professionally, you know that one-step writing is about as real as unicorns.

Just like a pro-writer, we started breaking down our prompts into a series of steps, just like a real writing brief.

We then asked the LLM to walk through the steps.

And then we realised we could ask the LLM to explain the work and ask us questions.

The tide is starting to turn.

7. The Embrace of Human-LLM Collaboration

Checkpoint: At this point, we are light-years away from our first fumbling attempts at using LLMs. Well, you and I are because we’re now on the same rollercoaster, holding hands and screaming together. But lots of people are still queuing up and watching us fly past with grimaces of both envy and fear.

We went from zero-shot asking the LLM to do something to teaching it, and then we started asking it to take our own guide and improve it. It was like handing over the keys to our creative kingdom, and the process looked something like this:

1. Brief the LLM on the task at hand.
2. Cast it in the role of guide, with you as the subject matter expert.
3. Ask it to walk you through the tasks and interact with you.
4. Engage in a Q&A session with your LLM collaborator.
5. Collaborate like you’re in a virtual writers’ room.

8. Handing the keys to the LLM with Chain of Thought

The latest evolution in our evolution of using — erm, now collaborating with — LLMs, is Chain of Thought (CoT) prompting.

CoT is like giving the LLM a mental GPS co-ordinate and watching it drive itself to the destination

Steve Chambers, Viewyonder

With OpenAI’s recent ChatGPT-o1 “Strawberry” model, this technique is becoming more accessible to non-AI folk.

🧠 Deep Dive: Chain of Thought prompting has shown significant improvements in tasks requiring multi-step reasoning. In a 2022 Google article, “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models“, researchers demonstrated that CoT prompting improved performance on math word problems by up to 40% compared to standard prompting techniques.

Imagine asking your old zero-shot questions, but now they come with all this evolutionary goodness baked right in. It’s like upgrading from a flip phone to the latest smartphone—same basic function, but oh so much more powerful!

Wrapping Up: The Future is Bright (and Slightly Terrifying)

As we stand here in September 2024, it’s mind-boggling to think about how far we’ve come in our journey with LLMs — both personally, and collectively.

From clumsy zero-shot prompts to sophisticated chain-of-thought reasoning, we’ve evolved alongside our LLM writing partners.

What’s next on the horizon? Will we be telepathically communicating with our LLM collaborators by 2025? Neuralink, anyone? Only time will tell. But one thing’s for sure: the world of content creation will never be the same. In fact, which non-manual area of business will remain untouched, I wonder?

So, fellow prompt engineers and LLM wranglers, keep experimenting, keep learning, and most importantly, keep that human touch in your work. After all, it’s the collaboration between human creativity and LLM capability that’s truly revolutionary.

Now, if you’ll excuse me, I need to go ask my LLM assistant to write me a sonnet about the existential dread of being replaced by LLM. You know, just for funsies.

---

Want to stay on the cutting edge of LLM content creation? Follow us on socials for daily tips and tricks! Have you discovered any mind-blowing prompt engineering techniques? Share them because we’re always eager to learn from our fellow LLM enthusiasts!

• LinkedIn
• X
• Facebook
• Mail