
## The Big Idea

LLM tools like ChatGPT, Copilot, and Cursor are great at helping you build, but they are a second brain - not a first one. Used well, they are an intern that speeds up your work. Used poorly, they are a liability. The trick? Give them context, direction, and oversight.

<a href="https://www.paddle.com/resources/mobile-app-monetization-guide?utm_medium=website&utm_source=partner&utm_campaign=core_conversions_fy2025_web_monetization_brand_awareness&utm_content=tiny-improvements" className="text-xs lg:text-xs xl:text-xs">
  Brought to you by Paddle
</a>

---

### AI Works Best When You Steer It

Most AI tools (Cursor, GitHub Copilot, ChatGPT) aren't great at figuring everything out on their own. They do best when you provide them with context. Although this guide is focused on software development with Cursor, the same principles apply to other tools and domains.

I have found that the best way to use LLMs is to treat them like an intern:

1. Tell them the goal. Spell out what you're building.
1. Give them the pieces. Feed them relevant files, not just one at a time.
1. Make a checklist. Break the work into steps you can track.

Here's an example of a recent set of instructions I gave to Cursor, while working on a new feature for Craftwork:

```
take a look at @timesheets_spec.rb - I need to produce an updated version of our timesheets API under the v2 namespace. It should use all of the same patterns as the other endpoints in @v2 do, and should use a new controller namespaced under api::v2 at `controllers/api/v2/timesheets_controller.rb`.

It needs a list view that can be filtered by a craftworker's id, by timesheet state (approved/inprogress/etc), and by production lead id.

The show view should return details for a single timesheet.

The update view should allow me to approve a timesheet, edit values, etc
```

That was enough to get me started working with Cursor, and it helped orient the LLM while building out the new feature. It had enough context for me to ask questions as I worked, and sped things up nicely.

### AI Will Get It Wrong (And That's Fine)

For any instructions you give 'em, LLMs will [confidently spit out answers](https://mikebifulco.com/posts/product-marketing-defy-expectations), regardless of how correct they are.

If you don't understand what they're doing, you're trusting a bad intern to run your company.

That's why you need to:
- Spot bad logic. If something seems off, it probably is.
- Debug iteratively. AI is great for refactoring-ask it to explain itself.
- Know when to take over. Sometimes, the best fix is writing the code yourself.

In the case of the Craftwork timesheets, I was able to spot a few issues, and after making some changes, steered the LLM to the right path (my typoes included -- LLMs actually deal with these really well!):

```
great, now can you llook at the controller and add any helpful missing scopes? You may want to peek at other controllers in the @v2 folder, as well as @timesheets_controller.rb

Alsothe .d.ts type definition declarations you made arne't needed - see @tsconfig.json and @javascript - they're likely in another folder already configured with typescript paths. If the components don't exist they'll need to be created
```

If you've ever onboarded a new teammate to a codebase, this kind of steering is not at all uncommon - you're helping them understand the codebase, and the context of the changes they're making.

### Tests Are More Important Than Ever

If you're letting AI write code, you better be _damn sure_ it works as expected. That means writing great tests before shipping anything.

- Test like a skeptic. Assume AI-generated code is flawed.
- Write failure cases. If AI wrote the function, you should write the test.
- Use AI to help debug. It's great at explaining why tests fail (sometimes).


## The Bottom Line

AI is a force multiplier - but only if you use it right. Treat it like an intern:

- Give it structure.
- Question everything.
- Make sure the work holds up.

Most importantly, remember that your great, big, juicy brain still need to do some work. Don't blindly trust the output of your copilot-of-choice - that's a hell of a lot to ask of an intern.


---

## Tests? What tests?

Most of us (hello, yes, me included) aren't great at writing tests -- but writing software tests is a skillset that is absolutely worth investing in. Here's some resources I've found helpful:

- (React) [Testing Library](https://testing-library.com/) provides documentation and set of tools for testing web applications, in a variety of frameworks. Whether you're a fan of react or svelte or vue or solidjs, this is a great resource.

- One of my favorite things about building with Ruby on Rails is the framework's first-class support for great testing. There are _so_ many thoughtful features for writing great tests for rails apps. Even if you're not a rails developer, it's worth giving their docs a look: [Rails Testing Guide](https://guides.rubyonrails.org/testing.html)

---

## Stop paying the app store tax.

Sell your app on the web with lower fees and full control. As a merchant of record, Paddle handles payments, taxes, and compliance - so you can focus on what matters: building your app.

Paddle acts as the reseller of your app, offering a solution similar to the App Store or Google Play Store but optimized for successfully selling your mobile app on the web, with significantly lower fees.

[Start selling smarter with Paddle.](https://www.paddle.com/resources/mobile-app-monetization-guide?utm_medium=website&utm_source=partner&utm_campaign=core_conversions_fy2025_web_monetization_brand_awareness&utm_content=tiny-improvements)

---

## Thinking about automation

I've spent a bunch of time working on small scripts to automate manual tasks lately - and I'm working on a couple quick video tutorial showing how you might use tools like Puppeteer, Faker, Skyvern and your copilot-of-choice to get things done more quickly than ever before.

If that sounds interesting, go chase me down on youtube at [@mikebifulco](https://www.youtube.com/@mikebifulco?sub_confirmation=1).


AI is your intern (not your boss)


## The Big Idea

Talking to AI is a lot like talking to people... except for all the ways that it isn't. It's a skill worth honing, and not a replacement for thinking.

---

From 2020 to 2022, I worked at Google.

I was on the Google Assistant team, in an era where "hey alexa" and "hey google" weren't new ideas. ChatGPT had not been released yet, and in general, the public was not aware of AI; "LLM" was far from a household term.

Funny enough, the dev product that my team was promoting was called "conversational actions" -- back then, "conversations" were predetermined - whether you realized it or not, you were following a scripted series of questions, and Google Assistant was a clever way to trigger actions.

In that time before LLMs took over, I noticed a funny pattern: people would talk to their assistant *like* a robot. They'd slow down, speak in weird, clipped phrases, almost like they were trying to sound like the assistant itself. It reminded me of how people adjust their speech when talking to someone who doesn't speak their language.

Fast forward to today: people are speaking to LLMs much more naturally. _Conversations_ with AI feel more convincing, but the reality is that the LLM is still just a tool. It's up to you to do the thinking.

Of course, LLMs are getting better all the time, and the tools will continue to get better. There's still a real skill in knowing *how* to talk to your LLM. Done well, it can be a powerful extension of your thinking. Done poorly, it's a waste of time.

After a bit of reflection, here's the pattern I've adopted for working with AI tools:

---

## Provide Context (More is Better)

Once upon a time - like, two years ago - LLMs had tiny context windows. They were only able to process a few paragraphs of information at a time. Now there's *loads* more headroom. Most modern models can handle essays, transcripts, full articles, and more. The more context you provide, the better the results.

Here's how I set mine up for success:

- **Tell it who you are.** Your background, goals, and what expertise you expect from it.
- **Define the task.** Are you brainstorming? Writing? Editing? Summarizing?
- **Give it sources.** If it can read the web, provide links. Otherwise, copy/paste key quotes or references.
- **Set guardrails.** Be explicit about what you *do* and *don't* want. Examples:
  - "Be concise."
  - "Avoid jargon."
  - "Don't mention XYZ."
  - "Your audience is nontechnical."

---

## Iterate and Customize

You'll get better results if you tweak your prompts while working, and you should save what works, and refine it over time.

- **Save what works.** If you get a great response, refine it into a reusable prompt.
- **Use different prompts for different goals.** You wouldn't write a social post the same way as an executive summary. You may want to have separate prompts for copywriting for social media, writing execustive summaries, or brainstorming on SEO keywords, for example.
- **Redo / Retry / Resubmit.** LLMs aren't deterministic: if you submit the same set of instructions multiple times, you'll get different results. If it misses the mark, tweak the input and try again.

I think it's also a good idea to regularly try different models and tools - what does Claude do better than Gemini? You'll never know if you don't experiment with the tools. Don't get lazy, and don't blind yourself to what's out there - it's all changing very quickly right now.

---

## Don't Get Lazy

An LLM is a tool - not a replacement for thinking.

I'll admit, I go into a bit of an existential spiral when I think about the future of communication - do we want to winnow down our existence to LLMs talking to each other? I certainly hope not. All the same, these tools are really useful for brainstorming, editing, and revising your writing, and they're not going away.

You're still the expert (ideally). It's your job to check facts, refine tone, and ensure the final product *sounds like you*. In my opinion, a big part of that work means that you should treat your LLMs like an assistant, not a crutch.

Edit the output - make it sound like you!

---

## TL;DR

1. More context = better output.
2. Experiment, tweak, and refine your prompts.
3. Never stop thinking -  you're still the one with the expertise.


---

## Your inbox shouldn't slow you down.

Superhuman's AI-powered features-Instant Summarize, Smart Replies, and Ask AI-help you fly through email, reclaiming at least 4 hours every week.

That's 4 extra hours to build, think, and focus on what actually matters.


📩 **Tiny Improvements readers get their [first month free](https://superhuman.com/ti?utm_medium=sponsor-callout&utm_source=mikebifulco-com).**

---

## Stick these in your brain's context

I'm a firm believer that we should always be learning. Don't rely on AI to do all your thinking. Here's a couple things I found this week that you might enjoy:

- [Ranking Paradoxes](https://www.youtube.com/watch?v=1vJsWKg2o1w) is a YouTube video from Chalk Talk, which goes discusses the complexity and counterintuitivity of philosophical and mathematical paradoxes.

- I'm still very much into 3D printing nerdery - and this recent video from Make Anything on [multi-texture prints](https://www.youtube.com/watch?v=zmZCA9jqGiI) shares a really cool new technique for making prints more visually interesting.


How I (Think I) Talk to AI


If there's one thing I appreciate about the internet, it's people who find themselves fixated on a topic, and go _really_ deep to research and find the answer to a question. It's a helpful reminder that the internet puts huge power at our fingertips, and the connections available to us are truly incredible. There's something really special about watching people become experts on a topic, and then watching experts do their thing.

Is hyperfixation a good thing? Maybe. Not always. But it's a good thing when it leads to something useful.

There is a similar pattern I see in many of the best engineers I know - a simple need or curiosity will lead to something much larger. Chasing down a solution is often just a series of small questions that need to be asked and answered in order, rolling up to a larger goal.

A few examples from the tech world:

- My friend Charlie found herself wondering about the jets she could see from her apartment window in Seattle. One thing lead to another, and she built a javascript app that gives her [real-time information](https://charliegerard.dev/blog/aircraft-radar-system-rtl-sdr-web-usb/) on the planes whizzing by her window from live, public radio signal broadcasts.
- Curiosity about modifying the generator command in Ruby on Rails led my teammate CJ deep into the weeds of the [RoR source code](https://www.youtube.com/watch?v=8Rnf_-vDuaU&) to figure out how it all works, and how to make it better - his work will likely find its way upstream into core parts of the rails ecosystem before long.

---

## My own hyperfixation

A friend recently recommended a _3-hour long podcast about Costco_ to me from the show Acquired - and while it is an incredible deep dive, it also requires a huge time commitment to get through. I listened to the show, and wanted to chat about it with my teammates, but I knew that it was unlikely they'd have the time to listen the entire 3-hour show.

So, I did what any good engineer would do, and I figured out a way to summarize the show's transcript for each of my teammates in a way that was personalized and relevant to them. I built a little tool that would take the transcript, and then use a series of natural language processing tools to summarize the transcript into a few key points.

In a short amount of time, I found myself learning about:

- **The Podcast RSS feed format**, which contains audio files and metadata about individual episodes. I used this to grab both a show mp3 and transcript to use for summarization.
- **OpenAI's Whisper API**, which can be used to quickly transcribe audio files into text.
- The **GPT-4 API Playground,** where I could play around with different summarization models to see which one worked best for my use case.
- **LLM Prompt writing** for summarization, and ways to ask the model to summarize an episode for a specific listener's perspective, like our head of growth or a software engineer. Clearly, these two personas will have different interests in a show about Costco -- so I wanted to make sure the summary was relevant to them.

I ended up with a tool that I could use to summarize any podcast episode, and I learned a ton along the way.

### Read summaries of the Costco episode generated by AI

Sounds cool, right? I published a few sample summaries of the episode in an [article on my site here](/posts/acquired-costco-podcast-llm-summaries) for you to check out.

Is this something you'd a small amount for? [Drop me a line](mailto:hello@mikebifulco.com) or shout at me on [threads](https://threads.net/@irreverentmike) - I'd love to hear your thoughts.

---

These are a few incredible examples of people who have gone above and beyond to pursue answers to _non-technical_ questions:

### The Mystery of the Bloomfield Bridge

[The Mystery of the Bloomfield Bridge](https://tylervigen.com/the-mystery-of-the-bloomfield-bridge) may be the single most interesting article I've read this year. It shows how a seemingly simple question about the built environment can uncover a long and winding history of civic action, economic change, and a bridge that doesn't seem to go anywhere.

### Finding Michaelsoft Binbows

YouTuber Nick Robinson's 42-minute long vlog titled [I flew to Japan to visit "Michaelsoft Binbows" in person](https://youtube.com/watch?v=QRIklga9IBQ) is actually a follow-up to a previous video where he tried to find the real-life inspiration for the Microsoft Binbows logo. Don't know what the hell that means? Neither did I, but I watched both videos and I'm glad I did. The internet is a weird place.

---

## Build your expertise with some of my favorite resources

### 📰 API Specs, Schemas and Standards

In the latest edition of the APIs You Won't Hate Newsletter, my pal Alex goes deep on [API Specs, Schemas and Standards](https://apisyouwonthate.com/newsletter/api-specs-schemas-and-standards/) - what's the difference between JSON Schema and JSON:API? Find out here.

### 🎙️ The future of HTTP - software engineering daily podcast

I did a super in-depth interview about [The Future of HTTP](https://softwareengineeringdaily.com/2023/10/11/the-future-of-http/) with Nick and Roman, two engineers from Nginx working on a standard that is one of the most basic building blocks of the web. I genuinely had to study up on HTTP basics for this one, and learned a ton. I hope you do too!

### 🧬 Your Resume Sucks

The gang over at Polywork liked my recent newsletter on resume writing so much that they [reposted it on their blog](https://blog.polywork.com/how-to-fix-your-resume/) for the world to see. I'm happy to see that it's resonating with people, and I hope it helps you land your next job!

<small>

_note: Cover photo for this newsletter by [C D-X](https://unsplash.com/@cdx2?utm_content=creditCopyText&utm_medium=referral&utm_source=unsplash) on [Unsplash](https://unsplash.com/photos/PDX_a_82obo?utm_content=creditCopyText&utm_medium=referral&utm_source=unsplash)_

</small>


Chase your interests to build something useful

Read summaries of the Costco episode generated by AI

Build your expertise with some of my favorite resources

🎙️ The future of HTTP - software engineering daily podcast


AI is _everywhere._

Chat GPT and interfaces like it have found themselves centerstage in tech culture. AI features have been added to search engines, word proccesing tools, image editors, audio editing software, and for crying out loud, even [Snapchat](https://www.theverge.com/2023/2/27/23614959/snapchat-my-ai-chatbot-chatgpt-openai-plus-subscription)! It's an exciting time, and there's lots to learn from the new frontier of AI-powered tools -- especially, I think, for people who don't work in heavily technical roles.

I've been writing about generative AI for a while (["Hey Influencers, be careful with ChatGPT](https://mikebifulco.com/posts/product-marketing-defy-expectations)") - and I'm really happy to see that it's finally starting to find its way into the mainstream. It's super interesting to see all of the creative ways that people are using it. Seeing people find ways to use ChatGPT through the lens of their own experiences, skills, and desires is truly fascinating.

Today, I want to talk about **using tools like ChatGPT to learn**.

<SponsoredSection
  sponsorName="RemoteMore"
  CTAtext="Get Your FREE Access"
  href="https://vpdae.com/redirect/mkq11hf6zu8b8q75uqbpwkmcsao"
>
  <div className="grid md:grid-cols-2 gap-4">
    <Image
      publicId="sponsors/remotemore/logo"
      alt="RemoteMore"
      height={250}
      width={250}
    />
    <div>

      ### Hiring developers made easy:

      - ✅ Search in 24,000+ remote developer profiles
      - ✅ All developers have passed our extensive screening
      - ✅ See their salary expectations upfront in their profiles

      Do you want to try our Hiring Platform for free?

    </div>

  </div>
</SponsoredSection>

One of the most common use cases I've seen for people using ChatGPT for the first time is - "tell me about `%some-interesting-thing%`". ChatGPT can distill complex topics into easily understandable summaries in a flash. It's thrilling -- seriously. But there's a tradeoff: ChatGPT's speed to respond can make it feel like we never need to think again. Got a question? Ask the robots. Curious about something? Ask the robots. Can't remember that one thing from high school physics? Ask the robots.

Offloading work from your brain is a good thing -- but making knowledge fully disposable isn't. Retaining _some_ of the things we learn is what helps us draw connections, foster creativity, and grow as people.

The [_Generation Effect_](https://en.wikipedia.org/wiki/Generation_effect) is the idea that you'll remember and understand stuff better when you put in the effort to do it yourself, rather than passively taking it in.

The thing is, when we take the time to experience and write about something, we're not just passively absorbing information. We're actively making connections, thinking critically, and making the material our own. It's the difference between watching a tutorial on YouTube about fixing your dryer, and then actually _fixing_ the damn thing yourself. Applied experience will do a lot more to cement learnings in your brainmeat™️.

When we engage with the material, we're more likely to remember it.

<Image
  publicId="newsletters/make-yourself-smarter-with-chatgpt/screenshot"
  alt="The secret to using chatGPT to LEARN"
  caption="I often use ChatGPT to spur ideas for writing"
/>

This is where tools like ChatGPT come in. These AI-powered dudes can be super helpful in getting our creative juices flowing. They can toss ideas back and forth with us, ask questions, and give us inspiration. By ChatGPT to kickstart our thinking, we're more likely to dive deep into a subject and learn more than if we were just scrolling through generated content on our own. That's a win.

So, the bottom line is that by getting hands-on and generating our own ideas, we stand to learn and retain so much more than if we're just skimming through content. And when we use ChatGPT to help us, we're taking advantage of the best of both worlds: the AI's ability to spark ideas and our own creative thinking to really make that knowledge stick.

For me, that means using AI tools to generate high-level bullets, and doing the work of expanding on those points and summarizing myself. When I'm doing research, the initial high-level points for my notes may start from ChatGPT,but in the process of summarizing, they often get revised to a completely different set based on my perspective. It's often enough to cause me to ask the bots for more _specific_ summaries, or a pivot from one topic to another.

For example -- this very newsletter started out as 5 bullets from chatGPT about something called the [Doorway Effect](https://en.wikipedia.org/wiki/Doorway_effect).

The simple act of going a level deeper to engage with the things generates can be _surprisingly_ productive. Give it a shot the next time you fire up ChatGPT - you might be surprised by the things you learn.

---

## Tucked away in my notes about taking notes

- 🗂️ The [Tickler File System](https://officedynamics.com/43-folders-method-paper-system/) is a physical system for organizing tasks that need to be completed on specific dates. It pre-dates digital note-taking, but is certainly a forebearer of more modern systems like [Getting Things Done](https://gettingthingsdone.com/). It's a surprisingly elegant system - 43 folders are used to represent tasks for each day of the year. I know -- I was surprised, too.
- 🎥 My brilliant better-half Natalie recently released a video about [ChatGPT in K-12 schools](https://www.youtube.com/watch?v=M9oOtOkPTzc&t=175s) on her YouTube channel, Primary Focus. The big idea: it's time to _lean in_ to AI for education, and not build walls around it. The revolution is here, my dudes.
- 🧘‍♂️ Axios published a great article on [Using ChatGPT for therapy](https://www.axios.com/2023/03/08/therapy-chatbots-chatgpt-ai). Therapy is a good thing, and _many_ people can't afford it. Chatbots are not a perfect solution by any means, but if they can help expand access to healthcare, that's got to be a good thing.

## Cache me outside

- 🎙️ My [interview with Guillermo Rauch from Vercel](https://softwareengineeringdaily.com/2023/03/15/building-user-experiences/) for Software Engineering Daily went live last week. Guillermo is clearly an extremely thoughtful and clever leader -- and I was lucky to get some time with him to talk shop.
- 🎤 I also interviewed [YC Founder Danny Sheridan](https://softwareengineeringdaily.com/2023/03/15/building-user-experiences/) about his new startup, [Fern](https://www.buildwithfern.com/). If you're building SDKs or client libraries for your company's API, you'll appreciate this one.
- 📖 As I've been prepping outlines for a quick course on note-taking, I was reminded that I've shared some of my notery on my site before. My article [Book Notes: Working in Public](https://mikebifulco.com/posts/working-in-public) is essentially a direct copy/paste of my notes on a great book from _Nadia Eghbal_ called [Working in Public: The Making and Maintenance of Open Source Software](https://amzn.to/3LAc0a6).


An idiot's guide to using ChatGPT to make yourself smarter

Tucked away in my notes about taking notes


It's no secret that I like playing with new tech, and lately there has been no shortage of AI tools making the news. By now, I've spent quite a bit of time testing out AI tools of various descriptions. It is encouraging to see consumer-facing AI products quickly becoming more useful.

Even on my own website ([mikebifulco.com](https://mikebifulco.com)), I've used AI for a pretty broad variety of tasks. This includes writing and publishing articles, and maintaining the code that keeps the site running. Some recent articles have cover art generated by Open AI's DALL-E. I bet you can probably even pick them out if you look closely! Until this moment not a single soul has asked me about where they came from. That's one point to the machines for their good work. 🤖

I also use GitHub's Copilot as a pair programmer when I'm working on personal projects. It's definitely not perfect, and I have my concerns about the way that GitHub/Microsoft are sourcing the code that Copilot uses to generate suggestions, but it's still a really useful tool. I find that it helps me avoid brain fatigue when I'm coding, even if it suggests solutions that are sometimes incomplete or incorrect - it's still a great way to get unstuck.

Of course, alongside good experiences, I've had my fair share of eye-opening and distinctly bad experiences with AI tools. In my favorite article ever, I wrote about a positively jaw-dropping moment I had when attempting to use one product to write blog posts. Sound interesting? I hope so - give it a read here: [Why no one talks about the best marketing campaign in history](https://mikebifulco.com/posts/product-marketing-defy-expectations).

It's important to point out that AI makes many people (myself included) nervous. We're anxious about what implications AI tools have on the future of our work, culture, and politics. Those concerns are valid - right now, I think one of the most important things you can do is get some experience using AI. This will help you understand its strengths and weaknesses in a practical way. It will also help you stay in touch with what's going on in the industry.

<Image
  publicId="newsletters/what-i-learned-from-using-ai/autumn"
  alt="A photo of a vast mountain landscape, showing blue-colored mountain tops extending into the distance"
  caption="A landscape from my recent trip to North Carolina's Blue Ridge Mountains."
/>

## AI tools you should try

These tools all use machine learning and AI to help you get things done. While most of them are not free, they all have some version of a trial or free tier that you can use to get a feel for them. If you haven't tried any of these, I recommend giving at least one a look - AI has come a **very** long way since the days of Goldeneye 64.

### No Code AI tools

📝 [Jasper AI](https://jasper.ai/) is a fantastic aid for people who write - it will help you write better emails, blog posts, and more. It's a great tool for people who want to write better, but get stuck on the details, or suffer from writer's block.

🎙 [Snipd](https://www.snipd.com/) is an incredible tool for people who listen to lots of podcasts. They use AI to automatically create an analyze transcripts of the shows you listen to. You can then use those transcripts to **snip** moments from a podcast, just like you'd bookmark a link to a webpage, or earmark a page in a book. I use it to take notes on podcasts that I listen to, and it's a great way to keep track of the things I want to remember.

🛋 [Interior AI](https://interiorai.com/) is a tool that lets you upload a photo of a room and it will suggest furniture that would fit in the room. It's pretty cool, and has been growing by leaps and bounds. Full disclosure: I happened across the product on twitter recently, and [have been helping its creator](https://twitter.com/irreverentmike/status/1579959127932968960) hone in on their pricing strategy using Stripe's tools. Apart from that, I'm not affiliated with, or paid by them in any way.

### For developers

✍🏼 [Open AI](https://openai.com/) are the creators of GPT-3, a language model that can generate text. It's pretty amazing, and you can try it out for yourself [here](https://beta.openai.com/).

🖼 [Stable Diffusion](https://github.com/CompVis/stable-diffusion) is an AI tool that you can run on your own computer to generate images from text. This one is for my more code savvy friends - it's written in python and not too challenging to use, but it does require some comfort with using a terminal.


What I learned from using AI in my work


One of the most famous ad campaigns in history is Yellow Tail wine's "Crazy Rooster" campaign. **It's so famous because it broke two cardinal rules of advertising:**

1. You don't want to be memorable for the wrong reasons.
2. You don't want to offend your customer base.

What made this campaign so effective was that it didn't just break these rules, but completely shattered them by doing the exact opposite of what conventional wisdom would tell you.

They built their entire brand around being offensive. They were known for making crazy claims like, "Yellow Tail tastes like s\*\*t." They put up billboards in airports saying, "If it wasn't for stupid people, we wouldn't have any customers." And they even used slogans like, "The only thing yellow about Yellow Tail is the color of our piss."  These things wouldn't actually make anybody want to buy Yellow Tail wine… Right?

What they found was that people loved it. And when people love something, they tell other people about it. The end result? A $500 million brand in Australia with no marketing spend. Why? Because they focused on the right things.

**Von Restorff Effect: Bonkers Marketing Campaigns Spawn Successful Brands**

<Colophon />

Ok... so _now..._ this is an important place to take a pause.

<Aside type="warning">

Up to this point, every word you read was written by AI. It's also **completely untrue**.

</Aside>

<Colophon />

I've been playing around with the idea of using AI to generate blog posts for my site, mostly out of curiosity. I wanted to see if AI could produce quality writing content, with the hope of comparing analytics for a few AI-generated articles to some that were written entirely on my own.

There are tools available online which let you feed a couple bullet points into a generator and select the type of writing you're looking for (article, email copy, bulleted lists, etc). You pick a few options and hit go, and in moments you've got some fresh new prose to use for whatever you please.

## Validating the "Crazy Rooster" campaign

After reading the generated article stub above, I wanted to find some images of the marketing campaign mentioned in the text, so off to the search engines I went. This, for me, was a truly jaw-dropping moment: There wasn't a single image I could find which mentions or references the "Crazy Rooster" campaign.

I dug deeper. I soon found that not only were there no images from the campaign, but I couldn't find any references to that ad campaign anywhere. Yellow Tail's site, marketing blogs, search engines, all returned _nothing._ I even searched for smaller fragments of text from the article above, and came back with nothing meaningful.

<Image
  publicId="posts/product-marketing-defy-expectations/crazy-rooster"
  alt="Search results for 'yellow tail crazy rooster' on DuckDuckGo showing nothing relevant"
  caption="A web search for the Crazy Rooster campaign"
/>

As far as I can tell, it never happened. It doesn't exist.

**In other words, AI had written a really compelling article which was _completely and utterly fabricated!_**

What I suspect really happened is this: the information I fed to the algorithm that this particular AI used weaved together partial anecdotes from a large number of publicly available external sources, creating a believable, compelling, yet non-existent story about a real brand's marketing strategy.

That is _fucking_ impressive... and kind of terrifying!

Friends, I'm here to tell you: the AI tools that are available right now can author content which is _shockingly_ good -- but **there is a massively important caveat** that we all need to understand about AI-powered writing:

<Aside type="info">

**Remember:** AI is just _guessing_ at the "Intelligence" part of its job (at least for now).

</Aside>

Can you imagine if I had accepted this as truth, and published it without surrounding context? The only reason I discovered this wasn't true was because I took the time to look up corroborating information - but it would be _super_ easy to skip that step, especially if you don't understand some of the fundamentals of AI.

## What is the Von Restorff Effect?

In principle, the Von Restorff Effect is simple: **given a group of similar choices, people tend to remember experiences or products which defy expectations** - the thing which stands out the most tends to be stickiest, and the more _memorable_ the unexpected experience is, the stickier something will be.

For an extremely simple example, let's say I gave you 10 seconds to memorize as many animals from this list as possible:

<div className="grid grid-cols-3 gap-2 sm:grid-cols-4 md:grid-cols-6">
  <Button className="inline-block">Wolf</Button>
  <Button className="inline-block">Beaver</Button>
  <Button className="inline-block">Camel</Button>
  <Button className="inline-block">Snail</Button>
  <Button className="inline-block">Llama</Button>
  <Button className="inline-block">Pig</Button>
  <Button className="inline-block">Mink</Button>
  <Button className="inline-block">Squirrel</Button>
  <Button className="inline-block bg-orange-500">
    <span>Owl</span>
  </Button>
  <Button className="inline-block">Giraffe</Button>
  <Button className="inline-block">Penguin</Button>
  <Button className="inline-block">Eagle</Button>
  <Button className="inline-block">Moth</Button>
  <Button className="inline-block">Sheep</Button>
  <Button className="inline-block">Camel</Button>
  <Button className="inline-block">Panda</Button>
  <Button className="inline-block">Koala</Button>
  <Button className="inline-block">Chameleon</Button>
</div>

I'd put money any arbitrary person remembering _Owl_ over some of the others mentioned in there.

This phenomenon can also be used to make _products_ more sticky by building _experiences_ which defy expectations - in any marketplace with significant competiton, creating a stand-out experience can give your product a massive advantage.

## How this post came to be

For articles on this site, I keep a list of topics that I think I want to write about in a big list. I slowly add to individual topics over time, and once I feel like there's enough information on a given topic to produce a meaningful article, I spend some time organizing thoughts from the note into prose, and then publish.

To generate this article, I grabbed one of the ideas from that list, which was this _Von Restorff effect,_ which is a [very real psychological phenomenon](https://en.wikipedia.org/wiki/Von_Restorff_effect) that caught my attention some time back.

I fed the topic and a few cliff notes into an AI tool to see what it could generate - really it was just, a few bullet points and the thesis statement for the article. I wanted to talk about how the Von Restorff Effect can be used in product marketing, to create something sticky; something that will be memorable.

And so I fed the tool and I clicked `go`. The article above is what came out of it. What I was really really impressed with initially was the _quality_ of the writing. No grammatical changes were needed, there was a beginning, a middle, and an end, and it put together a super compelling story that served the points I was hoping to make.

In short, I was really impressed... and then absolutely stunned when I realized what happened.

## Learning from this experience

So I came away with this experience with two important learnings:

<Aside type="note">

1. Ironically, **this experience was an amazing example of the Von Restorff Effect** for me. I froze in bewilderment the moment I realized what was going on, thoughts racing through my mind like I had reached the end of _Fight Club_ for the first time. It's a moment I won't soon forget.
2. This is a really important lesson in working with AI, and how powerful it is. If you're not aware of the capabilities _and limits_ of the tools you're using, you are entering dangerous territory.

</Aside>

## The Von Restorff Effect and product marketing

The nugget here for people building products is simple: **if you can create an experience that completely defies expectations, your product will be far stickier with your potential customers.**

## A caveat on AI tools

It's worth noting that I have intentionally omitted the names of the AI tool that led me to this experience. For me, this is because I'm struggling with whether or not it's responsible to point my readers to tools which can easily lead you into a false sense of certainty. _With great power,_ as they say...

Given that most of my audience are developers, I think it should suffice to say that I used a product based on [Open AI's GPT-3](https://openai.com/) - an extremely robust AI tool for generating language.

This is a story that I've struggled to tell because of how important I feel like this is.

I can tell you this much: I certainly won't be writing blog posts with AI any time soon, but I will be periodically checking in on this stuff to see how much better it's getting. I think there's an ethical responsibility to understand the limits and capabilities of AI content creation tools.

## AI and Code Generation with GitHub Copilot

If you're a software developer, you may have come across [GitHub Copilot](https://copilot.github.com/) - billed as "Your AI Pair Programmer", it's a freely available tool which will **generate code for you** with little prompting. You can do as little as writing a few comments, and copilot will generate everything from classes to algorithms to complex patterns login flows and cryptography.

<Image
  publicId="posts/product-marketing-defy-expectations/github-copilot"
  alt="Screenshot from a demo on GitHub Copilot's home page"
  caption="Copilot will write code with you.. for you... and alongside you."
/>

You _should_ really spend some time using Copilot if you haven't already. It's pretty amazing to write a few lines of pseudocode in comments and see what it can generate for you. It feels like magic.

### AI-generated code is still risky

But, just like when using AI to generate writing, be warned: GitHub Copilot isn't perfect - it's pulling from a massive amount of public code, and making guesses based on what you tell it to do. It won't necessarily write the world's most performant code, and in some cases it will generate code with serious flaws. Since its early release as a technical preview, there have been all kinds of problems reported with copilot, including one report which gained attention since it appeared that Copilot was generating code that included (other people's) valid API keys:

<Tweet align="center" tweetLink="pkell7/status/1411058236321681414" />

## A cautionary conclusion

As we introduce AI into more parts of our work and our lives, we need to be aware of the _risks_ of using it. As technologists, there's an ethical responsibility to communicate these risks to our non-techncial neighbors in ways that they can understand. It's my hope that this anecdote about trying to generate a blogpost with AI is something that can serve as a warning which you can explain to your aunt Judy, your boss, or the influencers in your feed who may not seem to understand the risks of using AI.

I'd love it if you shared this post with a friend - and if you've got thoughts on other experiments you'd like to see, let me know! Find me at [@irreverentmike](https://twitter.com/irreverentmike) on twitter, or [@mbifulco](https://github.com/mbifulco) on GitHub.

## More reading

- 📖 If you're interested in learning about AI fundamentals, I recommend reading <a href="https://www.amazon.com/gp/product/0465094279/ref=as_li_tl?ie=UTF8&camp=1789&creative=9325&creativeASIN=0465094279&linkCode=as2&tag=irreverentm02-20&linkId=3b9fba7a06bd067249ef768ce946c146" target="_blank" rel="noopener noreferrer">The Master Algorithm, by Pedro Domingos</a> (Amazon affiliate link). It's a great primer on various types of AI.
- 🎨 If you're into design and usability, check out my thoughts on why [you should never center paragraph text](https://mikebifulco.com/posts/dont-center-paragraph-text)


Hey, influencers: be careful with ChatGPT

The Von Restorff Effect and product marketing

AI and Code Generation with GitHub Copilot

#ai: 6 posts tagged

AI is your intern (not your boss)

How I (Think I) Talk to AI

Chase your interests to build something useful

An idiot's guide to using ChatGPT to make yourself smarter

What I learned from using AI in my work