Generative AI tools like ChatGPT are NOT one-stop solutions to all your content production challenges. However, using the right tools and prompts will radically cut the time required to produce effective material and supercharge your writing. 

But first, a health warning: Be careful, because Google says AI-generated content is against its guidelines. So use AI as an input to the production process and not a final output. 

GenAI speeds up every part of the production cycle. It’s structured as a conversation, it remembers what you’ve asked earlier, and like Google, is best played with until you feel you’ve got the best answer.  

It’s a bit like Google. The secret to unlocking the power of tools like ChatGPT is learning how to come up with smart prompts – the queries you type into the bot. Those who use Google professionally will have come across the use of structural language to get more out of it. GenAI is the same, just much more flexible and, therefore, complicated.

A word on prompts: The joy is that the tools work as a conversation using standard language. But they work even better if you use a few adjustments. The examples below make use of some important tweaks to simple verbal instructions.

My Journey with ChatGPT

I’ve been playing with ChatGPT for 15 months now and it’s been a fascinating journey of discovery. But at times, I have felt a gap between my experience of the tool and the enormous hype behind the underlying GenAI technology. I’m not a super-intensive user and maybe I should have worked this out quicker, but I thought I’d try to chronicle what has been a valuable learning exercise.

Stage 1: Search alternative

It’s November 2022, and I get caught up in the digital media frenzy and start using ChatGPT to see what the fuss is about.

I use it like I use Google.

I demonstrate it to a group of diplomats who want to improve the image of their country on social media and it generates a passable set of ideas – their minds are blown like mine has been.

Part of the attraction is its convenience – no need to comb your way through links on search results pages in the hope they will give you the answer you are looking for. You get one answer, very quickly. This is super cool.

But the novelty wears off when it becomes clear that it only gives me useful answers a third of the time. More often than not, I’m resorting to a time-consuming search.

I attempt to write a piece about India’s Agricultural Schools and ChatGPT generates a series of fictitious academic and media citations. It takes me hours of searching before I realise these don’t exist.

I stop using ChatGPT.

Stage 2: Structured prompts

Six months have passed, and I’ve been studying copywriting frameworks to help in my training work on online writing. I notice that digital marketing specialists on Twitter are promoting prompt frameworks that look very similar.

It dawns on me that my prompts have lacked structure – I haven’t been giving ChatGPT enough of a steer. I decide to give the tool another chance and home in on a couple of frameworks that seem to work:

RACE: Role, Action, Context, Expectation

• Role: Specify the role of the tool.
• Action: Detail what action is needed.
• Context: Provide relevant details of the situation.
• Expectation: Describe the expected outcome.

The specification of ‘role’ seems particularly powerful. I preface most roles with ‘expert’.

TRACE: Task, Request, Action, Context, Example

• Task: Define the specific task.
• Request: Describe what you are asking for.
• Action: State the action you need.
• Context: Provide the context or situation.
• Example: Give an example to illustrate your point.

The addition of an example makes answers much more useful.

These structured prompts are now giving me something I can use about 50% of the time. Enough to justify the extra time taken to devise the prompts, but still not quite the success rate that would justify the extreme hype around generative AI.

Stage 3: Devise my own framework

I combine my two favourite frameworks and, taking liberties with the sequencing of the acronym, come up with:

TRACER: Task, Request, Action, Context, Example, Role.

Disappointingly, this doesn’t make much difference to the usefulness of results, but it’s given me more confidence that I’m always providing enough context to give ChatGPT a fair chance.

Stage 4: Interactive prompts

I’m a year in now, and still feeling something is missing. There are new tools to play with – Bing’s Copilot, Google’s Bard, and lots of specialist tools. I try many of them.

ChatGPT has become harder to use because I’m having to spend longer providing all the context, and it’s still only working half the time. I’m often left feeling that I don’t know what I’m doing. It’s all still very unfamiliar and I don’t know what to ask.

I come across Brianna Brownell, a Canadian data scientist, who has what she calls the Flipped Interactive Pattern, and it’s a lightbulb moment:

From now on, I would like you to ask me questions to [do a specific task]. When you have enough information to [do the task], create [output you want].

From now on, when I ask a question, suggest a better version of the question to use that incorporates information specific to [use case] and ask me if I would like to use your question instead.

When I ask you a question, generate three additional questions that would help you give a more accurate answer. When I have answered the three questions, combine the answers to produce the final answers to my original question.

This is such a powerful approach – tell me the questions I should be asking you. This is cutting the amount of time I need to spend coming up with the perfect initial prompt and replaces it with an interactive conversation, which is actually quite good fun.

Now fewer than 20% of my efforts are going to waste, and it’s somehow more reassuring to get these interactive responses – as though you are both working to perfect the prompts.

Stage 5: Create my own ChatBot

Its 2024 now, there’s a ChatGPT store, and anyone who pays for premium access can create their own bot. I have a go and upload documents from my website and Google Docs to train ChatGPT to cover the topics I focus upon and to use my tone of voice.

This proves quite time-consuming. But it cuts right back on the time taken to draft new proposals, and refine the copy on my website. I occasionally see things that I’ve never written about, which is a little concerning.

By this point, I’ve now turned 180 degrees from treating GenAI as an alternative to search to using it as an intelligent assistant. It’s saving me several hours a week. I feel I have at least got some control over this new technology.

The Idiosyncracies of ChatGPT

Since it’s been around the longest, we know the most about ChatGPTs quirks and foibles. Here are some of the most important things to know:

Hallucinations: About 3% of responses include misinformation, according to AI search platform Vectara. Which is one of the reasons you can’t use it for direct output.

Compliments: ChatGPT gives better responses if you say nice things about it or offer a tip!

Word counts: Forget it! ChatGPT just doesn’t work that way. d count requests

1. How to find story ideas using GenAI tools like ChatGPT

If it’s time to publish something and it’s not obvious what to write about, then AI can be a great way to get ideation started. What it can’t do is answer the question, “Give me three story ideas with the greatest potential to go viral on social media” (although it will try!). You will have to do some thinking to start the ball rolling here.

“You are a health journalist. On the topic of reducing the impact of non-communicable diseases, give examples of the most successful policies.” 

“Give me three countries with innovative policies for reducing obesity”

“Give me the names of five celebrities suffering from chronic heart disease”

2. How to test your story’s audience appeal

Getting the ‘framing’ of your post is probably the most important step. ChatGPT won’t do this for you, but it will give you many options.

“The most successful hook formats in my sector are: This entity (X) is doing this thing (Y) to solve a problem like (Z); This entity (X) has a solution to the challenge of (Y); (X)  lessons from activity (Y) you need to know. Give me two more hook formats to try”

“I want to write about cutting deaths from noncommunicable diseases. What story angles would be of most appeal to a broad audience?”

3. How to write a story outline using AI

It’s tempting to rush straight to asking ChatGPT to write a post, but this step is best done separately – you need to run some sense-checks on whether it is really covering all the ground.

“Act as a BBC News journalist. Write a persuasive article outline about the variety of policies being used to reduce deaths from noncommunicable diseases”.

‘I am writing a blog post on X. Create a set of N bullet points I should cover.’

When you get the answer, add any additional points you feel should be covered.

4. How to turn a story outline into a draft

“Turn the outline into a blog post of 1500 words. Include subheadlines where appropriate.”

5. Experiment with writing styles

You may feel what comes back from ChatGPT is a little lifeless. Or lacks a good structure. You can remedy this by requesting a rewrite in a particular style.

Axios’ ‘Smart Brevity’ article: An “Axios story” is structured like this: “Why it matters” describes the key idea; “Driving the news” gives primary context; “What they are saying” lists quotes from people in bullet point format; “By the numbers” presents statistics in bullet point format; “Between the lines” offers secondary context; and “The bottom line” is a brief summary of why the story is important. Write an “Axios story” for this article: 

Buzzfeed ‘Listicle’: A “Buzzfeed listicle” story is structured like this: “Headline” consists of the number of items in the story and what they have in common; “Bullet-pointed section” has a subject line made up of the sequential number in the list and a summary headline”

Quartz Explainer: “Write a 1500-word Quartz explainer on ChatGpt. Include a summary of other technologies it is competing with, how fast the uptake has been, and what jobs might be at risk.”

Reuters News Story: “Reuters style” is as follows: the headline is ten words or fewer; the first paragraph is a summary of what is new and is 40 words or fewer; the second paragraph summarises the most important context; the third paragraph is a quote, subsequent paragraphs should be written in strict order of importance, every third paragraph should have a headline of no more than three words summarising the point of that paragraph. Create a “Reuters style” story for this:  

Rewrite this paragraph/sentence/post in the style of…

6. How to optimise headlines

Headlines are probably the words that are most important to the success of your piece. Ideally, they should be ‘workshopped’ with professional editors to ensure they work. But for most people, that’s just not possible. And the next best thing is AI.

“Give me three Buzzfeed headlines”

“Give me three Buzzfeed listicle headlines”

“Give me three Upworthy headlines”

“Give me three Sun Newspaper headlines”

“Give me three Guardian newspaper words that make use of puns”

7. Getting SEO right

“List ten keywords I can use for my blog about chatgpt and content marketing”

“Which of those keywords or phrases would be most cost-effective?” 

“Create a meta description of no more than 150 characters for this article: “

8. Making social posts pop

“Create a Twitter social post for this article:”

“What image would work best for the Twitter post?”

“What type of chart would work best for the Twitter post?”

“What hashtags would work best on Instagram?”

Further Reading

Learn Prompting goes into more of the mechanics of prompt engineering and has taken some effort to keep this simple for beginners.

Tribescaler is an app built on ChatGPT that specialises in generating story hooks and will also suggest Twitter posts.