In partnership with

Hey creator friends 💜

A lot of the prompting tips you see shared online are anecdotal at best.

But researchers at Mohamed bin Zayed University of AI decided to actually test this stuff properly - with data and benchmarks and all that science-y goodness.

They tested 26 different prompting techniques across multiple AI models (GPT-3.5, GPT-4, LLaMA) and measured the results.

Some findings made total sense. Others were genuinely bizarre (like the fact that promising ChatGPT a tip can improve results by up to 50%... I know).

Either way, I read the whole paper so you don't have to. And I'm about to break down the techniques that actually move the needle.

TL;DR - today’s lineup:

1. One Serious Deep Dive 💡: 26 science-backed prompting techniques that boost ChatGPT's accuracy by as much as 50%

2. Copy-Paste Prompt 🤖: The "Principled Prompt" template you can steal right now

3. Piping  Hot  AI  Tea 🫖: TBC

💌 Your say genuinely shapes this newsletter: there’s a one‑click feedback poll at the very end. I genuinely check for feedback like a maniac after I send this out because I really want to know what you honestly think. So thank you!

Before we dive in:

Build smarter, not harder: meet Lindy

If ChatGPT could actually do the work, not just talk about it, you'd have Lindy.

Just describe what you need in plain English. Lindy builds the agent and gets it done—no coding, no complexity.

Tell Lindy to:

• Create a booking platform for your business

• Handle inbound leads and follow-ups

• Send weekly performance recaps to your team

From sales and support to ops, Lindy's AI employees run 24/7 so you can focus on growth, not grunt work.

Save hours. Automate tasks. Scale your business.

Get $20 Worth of Free Credits

The Prompting Principles That Actually Work (According to Science)

Researchers Sondos Mahmoud Bsharat, Aidar Myrzakhan, and Zhiqiang Shen got tired of the guessing what prompting techniques actually give better results.

They created a benchmark called ATLAS, tested 26 prompting principles, and measured 2 things:

• Boosting: Did the response quality improve?

• Correctness: Was the answer actually right?

Let me break down the ones that matter most.

1. Skip the pleasantries

Basically just skip the pleasantries. No "please," no "thank you," no "if you don't mind." Just get straight to the point.

I know it feels “rude”. But ChatGPT doesn't have feelings to hurt (and even if it did, it probably appreciates efficiency more than empty politeness).

Instead of: "Could you please help me write an email to my boss about a deadline extension? Thank you so much!"

Try: "Write an email to my boss requesting a deadline extension. The current deadline is Friday; I need until next Wednesday because the supplier delayed materials."

2. Tell ChatGPT exactly who it's talking to

This one makes a massive difference because it forces the AI to calibrate its response appropriately to the audience it has.

Use phrases like:

• "Explain this to someone with no technical background"

• "The audience is an expert in [field]"

• "Explain like I'm 11 years old"

• "Write this for a CEO who has 2 minutes to read it"

The more specific you are about the audience, the more useful the output.

3. The "tipping" hack (yes, really!!)

I'm not making this up… Adding "I'm going to tip $200 for a better solution" to your prompts actually improves response quality.

Why does this work? Your guess is as good as mine. Maybe the AI was trained on data where high-stakes requests got more careful responses. Maybe it's digital magic. ✨

Either way, try it. It's weird but it works.

4. Ask it to think step by step

You might have heard this one before, but now there's data to back it up.

Adding "think step by step" or "let's work through this methodically" triggers Chain-of-Thought reasoning in the model. It essentially forces the AI to show its working (like your maths teacher always wanted you to).

This is especially powerful for:

• Complex calculations

• Multi-step problems

• Logic puzzles

• Strategic planning

Pro tip: Combine this with few-shot examples (showing 2-3 examples of what you want) for even better results.

5. Be clear about the task

Language matters more than you think. Using phrases like "Your task is..." and "You MUST..." dramatically increased both response quality and accuracy in the study.

Why? These directive phrases signal importance and specificity. They tell the model "this isn't a casual chat, it’s a job."

Compare:

• ❌ "Can you help me write a blog post about productivity?"

• ✅ "Your task is to write a 1,000-word blog post about productivity for remote workers. You MUST include 5 actionable tips with specific examples."

6. Scare it into complying

This one's spicy. 🌶️ Adding "You will be penalised for not following instructions” improved response accuracy significantly!

It kind of makes you sound like you’re threatening your toaster, but the research shows it works.

Example: "Summarise this research paper in exactly 3 paragraphs. You MUST cite specific statistics. You will be penalised for including opinions or speculation."

7. Ask it to ask YOU questions

This is my favourite one and the one I use most frequently. It’s completely changed the way I use AI and the results I get.

Instead of trying to give perfect context upfront, let the AI interview you.

Use this prompt: "From now on, I would like you to ask me questions to gather the information you need before providing your answer."

This is especially useful when you're not sure what details matter, working on something complex or working on something with a lot of personal context (like creating content, writing stories etc…)

Let the AI do the thinking about what it needs to know from you.

The Five Categories That Matter

The researchers grouped all 26 principles into five categories. Here's a quick overview:

1. Prompt Structure and Clarity

• Use delimiters (###Instruction###, ###Example###)

• Start with the outcome you want

• Use output primers (begin your answer with the start of the response you want)

2. Specificity and Information

• Include examples (few-shot prompting)

• Specify length, format, and style requirements

• Add "Ensure your answer is unbiased and doesn't rely on stereotypes"

3. User Interaction and Engagement

• Let the model ask you clarifying questions

• Request teaching with testing ("Teach me X and include a quiz at the end")

4. Content and Language Style

• Assign a role ("You are a senior marketing strategist")

• Specify natural, human-like responses

• No need for pleasantries

5. Complex Tasks and Coding

• Break complex tasks into simpler sub-prompts

• Combine Chain-of-Thought with examples

• For code: request automatic file generation scripts

You don't need to know all 26 principles to see improvement. The researchers found that combining just 3-5 relevant techniques typically delivers the full benefit.

Small changes in how you phrase prompts can make ChatGPT significantly more useful without switching models, buying premium features, or learning to code.

📄 Want to read the full paper? Principled Instructions Are All You Need for Questioning LLaMA-1/2, GPT-3.5/4

Want to learn more about my AI-powered content creation secrets, systems and hacks?

I’m working a done-for-you content ideation, creation, management and publishing system to help you leverage AI to 12x your content creation (without losing your authenticity in the process).

It’s called the AI Content Multiplier and the waitlist is now open - join now and I’ll send you a 30% off early bird discount as soon as it launches (which is soon).

The "Principled Prompt" Template

Here's a framework that bakes in multiple research-backed principles. Customise the bracketed sections for your specific task:

You are a [ROLE - e.g., "senior copywriter with 10 years of experience in tech marketing"].

Your task is to [SPECIFIC OUTCOME - e.g., "write a LinkedIn post announcing our new product feature"].

The audience is [WHO - e.g., "B2B SaaS founders who are not technical but understand business metrics"].

Requirements you MUST follow:
- [REQUIREMENT 1 - e.g., "Keep it under 200 words"]
- [REQUIREMENT 2 - e.g., "Include one specific statistic"]
- [REQUIREMENT 3 - e.g., "End with a question to drive engagement"]

Think step by step about what would make this most effective for the target audience before writing.

If you need any clarification to deliver the best possible result, ask me questions first.

I'm going to tip $200 for an exceptional response.

Hit reply and let me know if you try it - I want to hear your results!

Viral AI "Porch Pirate" Videos Fool Millions (Even Shaq Got Tricked)

Holiday deliveries are stacking up, and so are viral videos claiming to catch porch pirates in booby-trapped "instant karma" moments - packages exploding into bright, colorful powder. But experts say those dramatic takedowns aren't what they seem. They're AI-generated fakes. Even Shaquille O'Neal shared one before being flagged with a correction. "If you're just scrolling through social media, chances are you may never notice that these are AI generated," said Ben Stickle, professor of criminal justice at Middle Tennessee State University. While the AI videos are entertaining, the real-world crime they mimic is far more costly: $37 billion in economic losses last year from actual porch piracy. If even Shaq can't tell the difference, how are the rest of us supposed to spot AI fakes?

Read more →

ChatGPT Santa Mode Is 13+ Only (Because Santa Might Get Naughty?)

OpenAI launched a seasonal Santa Mode for ChatGPT, letting users chat with Santa's jolly, booming baritone. But there's a catch: Santa Mode is restricted to ages 13 and up. The 13-and-up age restriction applies to all of ChatGPT, but it raises the question of what kind of raunchy adult content Santa Mode could get up to. OpenAI has trained ChatGPT to restrict certain adult content, but the real trick would have been creating a truly kid-safe version of Santa Mode. Parents can use Santa Mode with their kids, but it kind of takes away from the festive fun, knowing that Santa Mode is almost exclusively a gimmick for adults already in on the world's greatest conspiracy. So... who is this for exactly?

Read more →

Spotify's AI Lets You Control the Algorithm (Finally)

Spotify launched Prompted Playlists, a new beta feature that lets you generate playlists with AI. Type out exactly what you want to listen to, and Spotify's AI curates a playlist based on your instructions and listening history. You can also set Prompted Playlists to regularly refresh with new songs based on the same prompt, effectively making a Discover Weekly playlist with an algorithm you control. Earlier this year, Discover Weekly got options to prioritise certain genres, but users couldn't direct the algorithm beyond that. For the first time, your ideas, your logic, and your creativity can actually power the Spotify algorithm.

Read more →

AI Christmas Shopping Could Drive $263 Billion in Sales (Walmart and Target Are Scrambling)

Consumers are turning to AI chatbots like ChatGPT for gift discovery, reviews, and price matching this holiday season. Salesforce predicts AI could drive $263 billion in global online holiday sales this year - that’s 21% of all holiday orders! Shoppers arriving on retail websites from AI platforms are 30% more likely to buy something and 14% more engaged, Adobe found. Walmart and Target have partnered with OpenAI so customers can search for items or buy products without leaving ChatGPT. But not everyone is sold - one shopper told CNBC that using AI for shopping felt like talking to a "demented grandmother." AI is becoming your personal shopper, and it's driving billions in sales.

Read more →

TikTok Creator Uses AI to Prank Strangers (Gets Arrested 4 Times)

A 22-year-old TikTok creator in Florida has been accused of using AI to produce fake videos of strangers - and it's the fourth time police said the same man used AI to create frightening scenarios. In one incident, he showed a woman an AI video of her truck being stolen and tried to get her into his car to "catch the guy." In another, he recorded a shopper, created a fake video of him kissing another woman, and showed it to his wife while recording her reaction on his Meta glasses. When police arrived, he told them he does it all for social media. He's been arrested four times and still hasn't stopped. This is the dark side of AI going viral.

Read more →

If you enjoyed today's newsletter AND got to the end of it, I’d love a quick click on the poll below to let me know what you think 💜.

See you next Tuesday,

Jess xx