“In today’s fast-paced world of business...” Have you seen that phrase used in a social media post or blog and thought “Did they use ChatGPT to write that?”
Or maybe, it was your post and it was your customer who just thought that about your content.
The emergence of Generative Artificial Intelligence (GenAI), specifically Large Language Models (LLMs) like ChatGPT, Claude, and others has been nothing short of revolutionary. These powerful tools have democratised content generation, offering even those without a background in writing the ability to produce blogs and social media content. However, as with any tool, the effectiveness of LLMs hinges not just on their capabilities, but on how we use them.
The difference between merely using LLMs and leveraging them to their fullest potential lies in the craft of prompting and on being an editor and not just a copy-paster.
In this blog I explain how to use GenAI and LLMs effectively for engaging content creation without making it obvious you've done so.
The Pitfalls of Zero-Shot Approaches and Weak Prompting
A common approach many people adopt is the zero-shot method—inputting a minimal prompt into an LLM and using its first output verbatim. LLMs produce good content from zero-shot but to quote leading AI expert, Andrew Ng, “[zero-shot] is akin to asking someone to compose an essay from start to finish, typing straight through with no backspacing allowed, and expecting a high-quality result.”
Furthermore, unless you provide good direction to the LLM, it will use whatever it thinks is good phrasing. Which often means the use of expressions like "In today's fast-paced world..." or words such as "elevate”. Telltale signs of unmodified LLM output.
This often results in content that, while coherent, lacks uniqueness and personal flair. Text can appear insipid, betraying a lack of genuine effort and insight.
Crafting Effective Prompts
The secret to transforming output from generic to compelling lies in crafting detailed, specific prompts. Ignore those click-bait infographics promising you “the only prompts you need to know”. Whilst there are many wrong ways to write prompts, there is certainly isn’t one perfect way. I've only seen one good "cheat sheet", it's based on detailed scientific research and also uncovered some curious quirks in LLMs.
Here are some important rules to use in forming a prompting strategy that works well for you.
Provide Context
Start by giving the LLM context about your intended audience and the desired outcome of your content. Are you addressing C-suite executives in the finance industry, or tech startup founders? Do you wish to inform, persuade, or entertain? Contextual prompts guide the LLM to tailor its tone, style, and substance, ensuring the output resonates with your audience. e.g. “The target audience is c-suite executives in mid-sized law firms who have a introductory understanding of the subject matter. On reading this blog they should feel like they’ve learnt more about the subject. It should also make them feel like our organisation is knowledgeable in this area.”
Assign a Persona
Assigning a persona to the LLM can drastically improve the results and alter the voice of your content, making it more relatable to your target audience. Whether you need the authoritative tone of an industry expert or the friendly advice of a peer, specifying this in your prompt can make all the difference.
This has two parts: 1) Assign a role to the LLM, such as: “You are an expert copywriter who works with specialist consultancies like ours [provide more info on who you are] to produce great content.” And, 2) Guide the LLM as to the Tone of Voice (ToV) for the content, e.g. “The post will be used on LinkedIn and needs to be professional in tone but remain light and easily readable by our C-suite audience.” You can also include a sample of your standard or preferred ToV and ask it to produce content in that style.
Specify Exclusions
To avoid the pitfalls of overused terms and phrases, provide an exclude list in your prompts. If certain words or clichés don't align with your brand's voice or the message you wish to convey, explicitly asking the LLM to avoid them can help maintain the authenticity of your content.
There's broadly two ways to do this. You can include it in your prompt, for example “do not use cheesy language and avoid terms like….” Or you can specify an exclude list: “you must not use any of the words or phrases in the Exclude list”. At the end of the prompt add a list like this: “## EXCLUDE elevate, fast-paced world of business…”.
Remember that LLMs are not perfect and there's a good chance it will still include words you don't want, so check the output.
Iterate: Embracing Few-Shot or Multi-Shot Prompting
Beyond crafting detailed prompts, adopting a few-shot or multi-shot approach can further refine the quality of LLM-generated content. Including the instruction “think this through step-by-step” has been shown to produce significantly better output.
You should also ask the LLM to take the content it just produced and iterate on it again, perhaps with specific guidance to change paragraph x or y. Multi-shot prompting has been demonstrated in certain circumstances to improve the quality of ChatGPT 3.5 output from 48% accuracy to 95%. And, of course, if you don't like what was produced, start again from scratch with a different version of the prompt.
LLMs are prone to verbosity. You can control this by specifying the length of the text, for example “max 3 short paragraphs for use on LinkedIn” or, “the blog must be no more than 500 words”. But also don’t be afraid to be blunt went iterating. “Make it shorter” will get the LLM to regenerate the copy it just produced in a shorter and more concise form. Note: LLMs notoriously struggle with numbers so 3 paragraphs might turn out as 4 and 500 words as 550. Again, check the output.
You are the Editor
Remember that this is supposed to be your content, so do an edit round (or rounds) yourself, then ask the LLM to critique your amended copy. This helps ensure the content is written in your voice and covers the points you want to make in the order you want to make them.
Furthermore, no matter how good your prompt is, LLMs often ignore your parts of your instructions, use words you've told it not to and invent facts. Be a critical editor and check the output throughly, particularly any quotations or references.
Five Steps to Great Content
Use the techniques described above in a five-step workflow:
- Started with a detailed prompt that includes context, a persona, and exclusions
- Ask the LLM to iterate on its output
- Edit the output yourself
- Have the LLM critique and/or polish the copy
- Review and publish it
Final thoughts: Master the Art of Content Refinement
Whilst LLMs produce fantastic content, at the current time they aren't perfect. However, the technology is improving rapidly and the use of cheesy words and phrases may be a fad that passes quickly. Certainly, ChatGPT 4 produces far better output than 3.5 and the recently released Claude by Anthropic creates remarkably superior content.
As we integrate these LLMs into our daily workflows, it's crucial to remember our role as editors, not mere copy-pasters. These models are a starting point—a raw material that we must shape and refine. By investing effort into detailed prompting and considering our content's impact from the audience's perspective, we can produce engaging, original, and insightful content.
In doing so we can delight our readers with great insights and ideas and avoid looking like we’re just trying to elevate our content in today’s fast-paced world of business.
Sample Prompts
Remember, there are no perfect prompts, the output from your prompt is just the starting point.
"I need a blog on how to create an information security policy. You are an expert copywriter who writes engaging content for business audiences and are working my company. We are an information security consultancy who works with mid-sized enterprises (>250 employees) who have an IT department but typically don't have a large dedicated security team. The target audience for this blog is C-level in those organisations who are interested in ensuring they have the right policies in place to protect their organisation. On reading the blog they should feel that they've learnt something about the subject and feel empowered to act on creating a policy. They should also see us as being experts in the subject who would be able to help and advise them but the blog must not be a sales pitch. The tone of voice should be professional, have a minimum level of jargon and be easily readable by our audience who are non-technical and are not experts in the subject. Use a maximum of 500 words, do not use any cheesy sounding terms or expressions. Our SEO keywords are [a,b,c] ensure these are included. Include an opener that sets the scene with stats on security breaches and the effect on organisations who don't take information security seriously. Provide a conclusion with a CTA to help them get started creating an information security policy. Think through the blog step-by-step before writing it. Where you use any stats or quotes you must cite all sources."
"Create a LinkedIn post for me to promote this press release by our company https://link.xx You are an expert in content marketing working for us, a software company who provides CRM software to large enterprises. Our target audience is C-suite and Heads of Marketing and Heads of Sales. The post needs to be short and to the point and engage the user and make them want to click through. Do not use the word "elevate" or similar or phrases like "in today's fast-paced world of business". Provide three versions of the post for me to review."
Postscript
If you've got this far you may well be asking "Did he use an LLM to create this blog?" Look, it's 2024 and I run an AI company, so of course I did.
I started with my traditional blogging approach of mind-mapping on a piece of paper what I'm trying to write, who the audience is and what the outcome should be. From this I formed my prompt and fed it into ChatGPT v4. I then iterated on it and heavily edited it before getting ChatGPT to give my content a final review. (Interestingly it disagreed with a particular tip I included about prompting despite there being independent research that proved it, I've since removed the tip while I investigate). I then asked it to generate 3 variations of the title I'd come up with.
What you've just read is approximately 50% LLM generated and 50% my own edits and additions to the copy. I've spent a reasonable amount of time on editing to get it to the final copy I'm happy with.
