Many users – rightfully - find themselves frustrated with generic, repetitive, or uninspired outputs from large language models (LLMs). The good news is that with the right techniques, you can coax these adorably eager-to-please machines into producing truly creative and polished content.

Understanding AI's Creative Potential

Before looking into specific tactics, understand that – despite what some still say - LLMs do not simply regurgitate pre-written text and are truly capable of creativity, just not nearly as good as the most creative humans. But having a draft generated in less than a minute which is in the ‘pretty good’ tier might save you several hours of editing or writer’s block and that still makes it a no-brainer for productivity. Remember that generative AI can be thought of as an eager to please intern who never gets tired but rarely gets things perfect – just because it’s not perfect doesn’t mean you shouldn’t use it at all!

The most advanced LLMs have been trained on vast amounts of data, allowing them to understand (artificial and limited understanding to be sure, but the results are what matters) and generate human-like text with impressive fluency. Infuriatingly however, their default settings often prioritise safe, middle-of-the-road responses. Many people assume that’s just all the AI is capable of, but it’s really not true. Give them a chance to reflect, review and revise and you’ll be amazed how much better they be pushed beyond the default, generic boundaries.

Be Explicit About Style

One of the most effective ways to enhance creativity is to provide a) clear instructions about the desired writing style, and b) examples of that writing style – in fact, LLMs are often much better at understanding from examples than from instructions. Instead of vague requests, offer concrete examples and detailed descriptions. By giving the AI a specific stylistic target, you're more likely to receive unique and tailored content.

Embrace Perplexity and Burstiness

Two key concepts in creating more human-like and creative text are perplexity and burstiness. Perplexity refers to the level of uncertainty or randomness in the text, while burstiness describes the variations in sentence structure and complexity. You may have noticed (even if only tacitly) certain sentence structure conventions from Chat GPT, e.g. “The noun verbs a noun, verbing a noun”, or to illustrate with a real example: “The intricate complexities of post-war Britain created multifaceted cultural phenomena, necessitating a blah blah blah”. Forget the fact it’s pure waffle saying nothing, that sentence structure dominates so much that it gets tiresome even when the content is useful.

Mixing things up is the way forward, and that means ‘burstiness’. Ask for a mix of short, punchy sentences interspersed with longer, more complex ones. Here's an example from one of this app’s pre-built templates, which has become popular on Reddit:

“.. Here is a good example of sentences with a high degree of perplexity and burstiness. Maintain a similar tone and writing style to this: "Premiere Pro has an attractive, flexible interface, and I'm a fan of the simplifying changes Adobe brought to it in the April 2022 update. The startup view helps you quickly get to projects you've been working on, start new projects, or search for Adobe Stock footage. The dark program window makes your clips the centre of attention. It now just has three main modes (in addition to the Home screen), for Import, Edit, and Export. A button or menu choice in Edit mode has a good selection of workspace layouts for Assembly, Editing, Colour, Export, and more. You can pull off any of the panels and float them wherever you want on your display(s).”

Avoid Overused LLM Expressions

AI models often fall back on certain phrases or structures that can make text feel robotic or clichéd. This means even words which were perfectly acceptable before now induce critical levels of cringe and even invite ridicule. ‘Delve’, ‘intricate’, ‘tapestry’, ‘meticulous’ – it’s sad that these otherwise perfectly nice words have become a sign of being generated by AI, but that’s where we are now. To combat this, the prompt can provide a list of no-no words or phrases. Many of the LLM Beefer Upper’s templates include these as standard, and of course once you add or create a template, it’s simple to edit and refine to match your preferred style.

Take advantage of the ‘temperature’ setting

Unfortunately the default commercial chat interfaces for Chat GPT and Claude don’t let you adjust the temperature of outputs. But this is a critically important feature that can dramatically change the style you get, because it controls the randomness of the AI's outputs – and when it comes to creativity, randomness (within reason of course) is what you want! A lower temperature (closer to 0) results in more focused and deterministic responses, which is great when you want to, say, generate an accurate FAQ based on a source document. But when you want to write something engaging, while a higher temperature (closer to 1) increases creativity and unpredictability.

LLM Beefer Upper has 3 settings for all templates, mimicking the Bing Chat / MS Copilot convention of ‘precise’, ‘normal’ and ‘creative’. This app currently has precise set to 0.3, normal 0.5 and creative 0.8. It may offer more refined options later, but things can get problematic at the extremes so for new users especially it’s worth keeping everything within the higher quality ranges.

For creative writing tasks (doesn’t have to be fiction of course, any writing style that you want to be more engaging is applicable), a temperature of 0.8 provides a good balance between coherence and creativity.

Provide Rich Context and Constraints

It may seem counterintuitive but giving the AI more constraints often results in more creative outputs. The key is to focus much more on positive rather than negative constraints. By providing specific style parameters and examples, you're challenging the model to think more laterally. The point can’t be overstated: lazy, generic prompts will get you lazy, generic outputs. The key value the LLM Beefer Upper brings is in giving you these editable templates which you can refine without having to hunt down and copy-paste every time. But don’t skimp on the initial effort, it really is worth it.

Encourage Metaphorical Thinking

Metaphors and analogies are the lifeblood of creative expression, and this is legitimately an area where advanced LLMs are seriously impressive out of the box. The very nature of how they were trained means they have a wildly complex - and imperfect of course – ‘all against all’ correlation matrix. You can ask for a story with totally absurd unrelated phenomena and it can still produce logical, plausible ways they can be brought together for a particular output. This is in fact one of the most entertaining things you can do with LLMs, it’s well worth experimenting getting the AI to think metaphorically by explicitly requesting comparisons or asking it to explain one concept in terms of another. For example:

"Describe the process of writing code as if it were cooking a complex dish. The target audience is a cat who understands English and is a master chef but hasn’t ever coded before. Use analogies and examples relevant to cuisine and the feline world to explain programming concepts to them."

Utilise Multi-Agent Approaches

Hands down the most powerful technique for enhancing AI creativity is to use a multi-agent approach, because the first output you get from the LLM is the equivalent of it blurting something out without property thinking about it. You can create a dialogue between different "personas" within the AI, each focusing on a specific aspect of the creative process. For example:

Agent 1: The Creative Writer (generates the initial content)

Agent 2: The Critic (analyses the content for areas of improvement)

Agent 3: The Editor (suggests specific enhancements)

Agent 4: The Polisher (produces the final, refined version)

This approach mimics the thoughtful, reflective, self-critical and iterative creative process used by human writers and can lead to significantly improved outputs.

Of course, implementing this manually is a huge pain. I know because that’s the reason I created this app in the first place. The LLM Beefer Upper automates the process of applying advanced multi-agent prompting techniques based on refined templates, making it a breeze to maximise the full creative potential that LLMs do truly have when they’re prompted well.

With LLM Beefer Upper, you can:

Create and save custom templates for different creative tasks

Use pre-built and custom templates (you have the option when creating a new template to get Claude Sonnet 3.5 to generate for you to fit LLM Beefer Upper's proven model) designed for maximum creativity

Automatically implement multi-agent refinement processes

Easily adjust settings like temperature

Provide rich context and constraints without repetitive prompting

By streamlining these advanced techniques, the app allows you to focus on your creative vision while the AI does the heavy lifting. As always, it will never be as good as your own best effort – but if you’re got deadlines left right and centre and the idea of starting the new draft is too daunting, you can get a draft done in seconds that does 80% of the leg work for you. At 75c/58p, it’s a no-brainer given the typical knowledge worker’s hourly rate!

***

Connect with me on Linkedin and follow me on Twitter.