You're the secret sauce: writing original content with AI

February 25, 2024
Michael Taylor

I hate to say it, but those AI bros on Twitter are right: 99% of people use ChatGPT wrong!. The most popular use case is content writing, as evidenced by the early growth of companies like Jasper and Copy.ai. Most people just give the AI a single prompt then use the output or complain the results are too generic.

Those who get serious about AI quickly graduate from “Write me a blog post on {topic}”, using prompt engineering to arrive at something that looks like this:

```Create a blog post about “{blogPostTopic}”. Write it in a “{tone}” tone. Use transition words. Use active voice. Write over 1000 words. Use very creative titles for the blog post. Add a title for each section. Ensure there are a minimum of 9 sections. Each section should have a minimum of two paragraphs. Include the following keywords: “{keywords}”. Create a good slug for this post and a meta description with a maximum of 100 words and add it to the end of the blog post.```

source

Once the post is written, a human editor takes over, and fixes any hallucinations (where the AI confidently makes something up), as well as rewriting it in the house style and attempting to make it more human and interesting. AI generated content tends to be a little dull and overly comprehensive, which is what you get when you train something on the average of the internet and beat it into submission with reinforcement learning.

This is exactly the wrong thing to do. They have it backwards.

ChatGPT is a helpful assistant, but it isn’t the star of the show. It isn’t human, so it doesn’t have life experiences that we would relate to. It’s too perfect. It doesn’t make mistakes. It can’t be the hero of the story, because it doesn’t have any flaws to overcome to create the drama. It knows the sum total of all human knowledge, but not how to give it meaning. It can imitate great art, but you need to tell it what to imitate.

"In order to write about life first you must live it." - Ernest Hemingway

Unopinionated content may as well not be written. When everyone is writing about the same keywords, quoting the same sources, generating the same copycat content, the search results pages become undifferentiated and unusable. When everything looks the same, the articles that stand out will be the ones that actually get read (by humans).

Google calls this ‘information gain’ in a patent they filed in 2020, with the aim of rewarding content for being different. Every search results page is already personalized and soon enough will be processed by AIs more than humans. Neither AI nor humans want to see the same thing 10 times, they want to see 10 novel things, so they can reason about what is better for the current situation they find themselves in. Aim to be a net contributor to the training data, not just a consumer of it.

Here’s how the process works:

  1. Browse model summary – use an AI model capable of browsing the web to look up the topic and summarize it.
  2. Conduct an AI interview – get ChatGPT to interview you about your experience to find unique opinions and insights.
  3. Generate an outline – prompt to make an outline that incorporates the summary and your interview answers, with title options.
  4. Write one section at a time – loop through the outline and generate each section, providing feedback if needed along the way.
  5. Rewrite in the right style – define a style guide and ask the AI to do a rewrite to produce the final article.

These stages don’t occur all in the same prompt, they’re chained together in a systematic way to deliver better end results – AI chaining is one of the 5 principles of prompting, and is used in most production AI systems. Each part of this sequence is designed to minimize hallucinations while maximizing creativity, and make original content worth reading without taking up too much of your time.

Let’s go through these one by one.

1. Browse model summary

We start simple, with asking the AI to look up the topic. It’s essential to have a model with access to the internet, so if you don’t have the “Browse with Bing” plugin for ChatGPT, use the Bing chat mode, or Google’s Bard. This is the number one thing you can do to avoid hallucination and be up to date with the latest on the topic.

Prompt:

```Please look up {topic} and provide a summary:```

There are three reasons we start with this:

  • It stops the AI hallucinating on topics it doesn’t know about.
  • Without it, the post tends to not be comprehensive enough and can be outdated.
  • It acts as a prompt to you, to get you thinking about the topic.

This doesn’t have to be done in the same chat window as the other prompts, or even using the same model. Just remember to copy and paste across for stage 2.

2. Conduct an AI interview

This is where the magic happens: you get ChatGPT to interview you. It’s actually pretty good at coming up with interview questions, especially when it has a recent summary of the topic to go from. You want the full list of questions at once, usually five is enough, otherwise it loses coherence and goes off on a tangent when doing questions one at a time.

Prompt:

```Ask me five interview questions on {topic}. Your goal is to discover any unique opinions, experiences, or insights not covered in the summary:```

I had a lot of these parts in place for a long time, but only recently cracked it by adding the interview stage. It was the missing piece, which is now reliably giving me good quality, creative writing that passes the Turing test, avoids AI detectors, with minimal or no editing needed. Rather than expecting ChatGPT to write something original, you provide the secret sauce. You’re the one the audience wants to hear from, ChatGPT is just your ghostwriter.

Note: if you don’t have anything interesting to say, find someone who does. Tap your network or find someone at your company who is an expert on the topic. You can still use ChatGPT for all the heavy lifting, just get your expert to record a loom video and send you the transcript. I've heard from a few people who've used this system and in the process realized they had nothing useful to say on a topic. And that's fine! You can use this as an indication of what not to do, as Tom Roach advises:

3. Generate an outline

Most prolific prompters have stumbled upon this trick: rather than asking for a blog post, ask for an outline and ask it to fill that in. Generating the outline ahead of time helps make the whole post coherent, and ensures you don’t get duplication across sections. This is based on the least to most technique, which is useful for other complex tasks involving AI.

Prompt:

```Based on my answers and the summary, generate an outline for a blog article on {topic}, with 10 very creative title options:```

At this stage it’s important to give feedback and make edits if necessary. You can drop sections or add them just by chatting, or copy and paste elsewhere to edit before pasting your final version back into chat. If you’ve done some keyword research on the topic now is a good time to drop that in, to make sure primary and secondary keywords are represented in the headings.

This is also the time when you choose your title, which is the one thing I find AI never gets right. However I do find that asking it to generate 10 titles for me does give me enough creative inspiration for me to write the final title myself. Make sure to tell ChatGPT what title you chose in the end, so it can match the writing to it.

4. Write one section at a time

One weakness with AI models that anyone who uses AI to write content at scale will have run into, is that it’s woefully bad at knowing how long an article is. It regularly doesn’t follow the number of words you prompt it with, and will write a short article with no real substance. Other times it writes too much, and has to be cut down. The antidote is writing each section of the outline one at a time, because it forces a kind of symmetry and consistency. 

Prompt:

```Write one section of the blog article at a time, with a minimum of two paragraphs. Incorporate my feedback before continuing to the next section:```

Typically I find that it needs the context in the interview and summary to write these sections correctly, so if you do start a new chat session be sure to copy and paste those across. If it makes mistakes, tell it, and it’s usually pretty good at learning from them. For example if a section is too short, correct it, and usually the following sections are longer afterwards. One issue I’ve seen is a tendency to hallucinate new sections as it gets lower down the context window, so always check each new section name back to the original outline.

5. Rewrite in the right style

To enact this last prompt, you have to know what style you want to write in. ChatGPT does know about the styles of prominent bloggers, brands, and celebrities, so you can ask it to emulate Seth Godin, the HubSpot blog, or Malcolm Gladwell. One strategy for getting something familiar that works but is slightly more original, is “unbundling”, a term coined by Bakz T. Future. Jump into a different chat window and run this:

Prompt:

```Briefly describe the writing style of Harry Dry from MarketingExamples.com in bullet points:```

From here you can remix the style, keeping the bullet points you agree with and dropping or replacing those you don’t. If you want to build an in-house style guide, simply copy and paste 5-10 of your best blog posts into ChatGPT and ask it to describe your style. If your posts are lengthy, you may need to use Anthropic’s Claude model for this, as it has a 100k token window, or about 75,000 words. Another neat trick is to use this when writing guest posts, to align your writing to what has already been published on the site.

Now you have your style, you can go back to your other chat window with the article in it, and run this simple prompt:

Prompt:

```Rewrite the article in the following style:
{ style }```

The style I used was as follows:

  • Concise and to the point
  • Professional and knowledgeable tone
  • Relatable and accessible language
  • Uses storytelling to convey ideas
  • Relies on examples and case studies
  • Mixes personal anecdotes with industry insights
  • Provides actionable advice and tips
  • Uses subheadings and bullet points for easy readability

Note: this is probably the only section you really need GPT-4 for. GPT-3.5-Turbo (which powers ChatGPT) is unreliable at following instructions for role playing, whereas GPT-4 is smarter and was specifically trained to solve this issue, and the quality is worth paying for by signing up for ChatGPT plus.

Bonus: Meta-prompting for images

Here I spent a bunch of time writing these prompts, but just like any text ChatGPT is pretty good at writing prompts too. When you get an AI to write prompts for you, it's called meta-prompting, and it works particularly well when it comes to creating a custom image for your blog, by asking it for a prompt you can plug into DALL-E (I actually use Midjourney, but I just wager ChatGPT knows more about OpenAI's own tools).

Prompt:

```Describe an image that would go well at the top of this article```

```turn this into a prompt for DALL-E```

This was a big insight when my co-founder on Vexpower showed me it. I just never thought about getting ChatGPT to write my prompts for me, I was already greatful it was doing everything else! If you wanted to learn more about this technique James covers it in our Udemy Course: "The Complete Prompt Engineering for AI Bootcamp (2023)".

Conclusion

I’ve found this system is capable of writing content that would normally take me more than a day or two in less than an hour. I have infinitely more stories and insights than I have time to write up, so now I can give my audience more of what it wants, and leave the boring, time-consuming parts to ChatGPT.

If you want to see the finished result, which took really minimal edits (mostly the bullet points section), here's my post on value-based pricing in agencies: https://www.saxifrage.xyz/post/value-based-pricing

Comparing the final text to what the AI outputted, a human editor (me) deleted 824 characters (8.54%) and inserted 1,016 (10.53%) of the text. Most of this was in the bullet point section in the last section on mitigating risks, which I had to rework.

There are plenty of techniques out there for using ChatGPT to write blog content for you, but I haven’t seen many like this. I've been battle-testing this for months, doing parts of my work with a mixed bag of these techniques. It was only recently that it all came together. The big unlock was reversing roles and getting the AI to interview me as the expert. Now I know it works, it feels obvious: of course things it writes aren’t meaningful if there’s no human experience to relate it back to. I actually hate editing, so I prefer ChatGPT to do that for me, and all I have to do is answer a few questions. I’ll keep refining this, and I hope you find success with it. If you find anything that works better, tweet at me: @hammer_mt