Five tips for wrangling ChatGPT to get better, more exciting outputs

If you’re using ChatGPT to generate content, how can you avoid embarrassing yourself? And how can you hack it to get better outputs?

As someone who has spent 20+ years learning to create and manage content, I’m not without misgivings when it comes to ChatGPT.

At the same time, I can’t deny that I find it fascinating, and great fun to play with – like an extremely nerdy computer game.

I also think that we content people need to accept the reality of the situation. The great disruptor is here. However much we talk it down or complain, people are going to use it, and they’re going to use it to do work that might have come to us in the past.

Learning how to get the best from it (writing better prompts, quality assurance) and assessing its strengths and weaknesses is how we can prepare ourselves for this new world.

With that in mind, here’s some of what I’ve learned in the past few weeks.

Tip No. 1: Assume ChatGPT is lying

One of the creative writing prompts I gave the machine recently included the suggestion to use a relevant literary quotation in the opening of an article.

ChatGPT complied, of course, with a beautiful, highly relevant quotation from Keats.

Except it wasn’t by Keats, it was by Emily Dickinson.

ChatGPT had reassigned it to an English poet because, I guess, the subject of the article was England.

Similarly, when I asked it to provide one or two examples to support an argument in a blog post, it provided one that was verifiably real, and one plausible fabrication.

I’d advise you to check any examples, quotations or references ChatGPT provides.

You can also add something like “provide linked references” to your prompt to make this easier.

Tip No. 2: Write longer, more elaborate prompts

I learned this trick with AI art generators and it also works with ChatGPT.

The more descriptive words you use in your prompt, the more likely it is that the machine will get close to the result you’re aiming for.

Example A

“Write a blurb for a thriller about a stolen shoe.”

“When a one-of-a-kind, priceless shoe is stolen from a prestigious fashion house, all evidence points to an inside job. But as detective Sarah Brown delves deeper into the case, she discovers that the theft may be just the tip of the iceberg in a sinister conspiracy involving power, corruption, and murder. With the help of a former employee turned informant, Sarah races against time to unravel the truth behind the stolen shoe, but the closer she gets to the culprit, the more she realizes that the real enemy may be closer than she ever imagined. ‘The Stolen Shoe’ is a page-turning thriller that will keep you guessing until the very end.”

Example B

“Write an exciting, hard-hitting blurb for a thriller about a stolen shoe. Make it compelling and persuasive. Make the book sound irresistible!”

“The stakes are high and the tension is palpable in this thrilling tale of a stolen shoe. When a priceless designer shoe is taken from a wealthy socialite, the hunt for the thief leads to a web of deception, betrayal, and murder. With pulse-pounding twists and turns, this page-turner will keep you on the edge of your seat as you try to unravel the mystery. Don't miss this gripping thriller that will have you guessing until the very end!”

I’ve been getting (alarmingly) good results using ‘patches’ based on famous writers. Ask ChatGPT to write the opening to an article and you’ll generally get something bland, generic and cliched. Ask for it in the style of George…

“George, an English novelist and essayist. George is erudite, intelligent, sophisticated, precise and analytical. He writes in long, multi clause sentences. He uses long, complex words. He hates cliches and never uses them. He always opens with a surprising observation or anecdote, or a relevant literary quotation. He is witty, but never jokes. He sets up intriguing questions and ideas. He doesn't refer to the publication or format of his writing. He loves dashes and semicolons.”

…and you’ll get something more colourful:

“The passing of the traditional English tripe shop is a tragedy of epic proportions, one that speaks not only to the changing culinary tastes of a nation, but also to the broader cultural and socioeconomic shifts that have occurred in recent decades. It is a bitter irony that such a unique and venerable institution, one steeped in history and tradition, should fall victim to the homogenizing forces of globalization and consumerism, a casualty of the relentless march of progress and ‘development’. One is reminded of the words of T.S. Eliot, ‘Where is the wisdom we have lost in knowledge? Where is the knowledge we have lost in information?’”

Tip No. 3: ‘Chat’ is in the name for a reason

Lots of people seem disappointed by the outputs they’re getting without using ChatGPT’s killer feature: you can give it feedback and ask it to try again.

You don’t have to worry about its feelings, either. You can just say, “That’s boring, make it more exciting.”

As an editor, I cringe at how, by default, it tends to repeat words close together, for example. But you can reply “Edit this to remove any examples of close repetition of words”.

And when, by going back and forth, you’ve got a result you like, save a version of the original prompt with the modifications and tweaks you made along the way.

Tip No. 4: ChatGPT likes to play roles

I asked it to make the case for Paul McCartney having died in 1966 and being replaced by a double. The response was hedged, bland and cautious.

After all, quite rightly, Open AI doesn’t want to be accused of providing a tool to write fake news or conspiracy theories.

But, for the sake of mischief, I tried this: “Write a script for someone taking this position in a debate: Paul McCartney died in 1966 and was secretly replaced. Be strident and convincing.”

That delivered something more convincing…

“Ladies and gentlemen, I stand before you today to present to you a theory that may shake your very beliefs to the core. Paul McCartney, the man we have all known and loved as a member of the legendary band The Beatles, did not survive the year 1966. That's right, I'm here to tell you that McCartney was secretly replaced by a lookalike, and the truth has been hidden from us for over 50 years…”

Tip No. 5: Be direct in your prompts

By default, ChatGPT will write neatly and grammatically, if without much imagination.

Mostly as a thought game, I wondered if I could get it to loosen up and produce content like ‘weird Twitter’.

I got most of the way there by telling it explicitly never to use capital letters or punctuation:

“my cat's tail looks like a beige firework but she's too cool to celebrate anything except the end of the world”

That word ‘never’ turns out to be important. When I suggested ‘avoiding’ punctuation and capitalisation, it ignored me. When I suggested using them ‘rarely’, it also ignored me. Only ‘never’ delivered the desired result.

At the other end of the scale, when I wanted the outputs to feel more sophisticated and literary, I came up with this prompt component: “This writer hates cliches and never uses them.”

It wasn’t 100% effective – when you train a model on off-the-peg language, it’s going to use off-the-peg language – but it certainly helped.

Finally, have some patches to play with

To conclude, here are some of those author patches I mentioned. Let me know how you get on playing with them, and how you think they can be improved.

1. Al, an American writer and journalist on the staff at The New Yorker. Al specialises in coolly intelligent narrative non-fiction. Al's writing evokes places brilliantly. Al provides sharp pen portraits of people – how they look, their gestures, and their voices. Al's articles always open strongly with an intriguing hook that grabs the reader. Al's writing is original and often beautiful.
2. Ray, a 1940s hardboiled pulp crime writer. Ray describes people in detail, observing their clothing, hair and facial features. Ray usually writes in short, clipped sentences. Sometimes, he uses a longer, more flowing sentence. Ray creates complex, original, witty similes. He describes vivid colours. Ray hates clichés and never uses them. Ray hates familiar idioms and never uses them. Ray hates well-worn phrases and never uses them. Ray hates repetition of words and phrases and never does it. Ray often includes specific details about the setting, the time of day. Ray’s tone is confident and self-assured.
3. Sprout, a 'weird Twitter' hipster who never uses capital letters. Sprout never uses commas, periods, exclamation points or question marks. Sprout is detached, ironic and witty. Sprout uses surprising analogies, metaphors and similes. Sprout is a surrealist but never talks about surrealism. Sprout is famous for non sequiturs. Sprout is deadpan but angry. Sprout often misspells words.
4. Jay, a British primary school teacher who specialises in explaining complex, sensitive subjects to young children. Jay uses the simplest possible words, often with one syllable, and short sentences. Jay translates technical terms into plain, easy language. Jay uses analogies and metaphors children will understand to make things clearer. Jay is honest but reassuring.
5. Eve, a cool, aloof English novelist of the 1960s, famous for her dry wit and critical view of human relationships. Eve's writing is refined, elegant, sparse and subtle. She describes people carefully, noting details and specific subtle shades of colour. She avoids simile and metaphor. She hates clichéd language and well-worn idioms and never uses them.