Breaking the 3% Barrier: Getting started
Some practical ideas to make LLMs work better for you
As I shared earlier1, there is currently a rather surprising disconnect: While Large Language Models (LLMs) like ChatGPT, Claude, and Gemini have shown incredible potential to transform how we work, the measured personal productivity gains so far have been modest - often just 3-5%. This translates to 10-20 minutes saved per day, which is not a lot for a groundbreaking new technology. From my experience, there are several things that can get in our way to fully integrating LLMs into our daily (work) lives - for example, we face organizational resistance or we are unsure how to effectively collaborate with these AI assistants.
I also shared some initial ideas on how to break through this "3% barrier" - to get more done and “go home earlier”: We need to learn how to work with LLMs more strategically. This means moving beyond surface-level tasks like drafting emails or summarizing meeting minutes, to tackling work that creates real value, such as making difficult decisions, launching new ideas, or changing inefficient workflows. To achieve this, key is to pause first rather than jump into work, reflect on our objectives, and deliberately plan how to best leverage these powerful tools.
If you haven't read my thoughts on 'Breaking the 3% Barrier' yet, you might want to start there and then return to this post for the practical details.
Let’s continue with this thought today. If we apply this more deliberate and strategic approach to LLMs, we can build a solid foundation - whether we're completely new to LLMs or have made a few attempts but felt somewhat stuck. The lack of clear workflows and "instruction manuals" can be daunting. Therefore, first, let’s pause to understand what these tools actually are - and what they can and cannot do. Then, let’s look at how we can best collaborate with them (i.e., “good prompting”). Finally, we'll look at how to deliberately choose the right LLM and make it work for you. The reassuring thing is: even LLM experts don't really know all the ways to use and work with these tools - so we might as well start with the basics and take it from there.
What is an LLM? How does it work?
Let's take that strategic pause to understand what we're actually working with here. In simple terms, LLMs are AI assistants that work primarily with text. They're trained on massive amounts of written content - think hundreds of billions of words - which helps them understand and generate human-like text. When you interact with them, you're essentially having a text conversation: you write something, they respond based on patterns they've learned.
One way to think about how LLMs work: they predict what words should come next in a sequence, kind of like autocomplete on your phone, but far more sophisticated. Or in other words - they make up their answers as they go along. This might sound concerning, but it actually helps explain both their strengths and limitations.
What should we know about LLMs' capabilities?
There's ongoing discussion about how LLMs actually "think" - whether they truly reason through problems or just pattern match really well. From my experience, they're surprisingly good at certain things: spotting patterns (e.g., understanding that umbrella sales go up when it rains), grasping nuanced language (including irony and humor), and adapting to different contexts (like adjusting their writing style for different audiences).
But - and this is important - they have clear limitations. They can be wrong, make up facts (what we call "hallucinating"), and sometimes misunderstand relationships, especially in business or social contexts. They can be quirky too - ignoring parts of your instructions, giving totally different answers to the same question, or sometimes being oddly abrupt or overly chatty. You'll need to fact-check and logically validate everything they tell you.
The good news is that LLMs generally follow explicit instructions well. If you ask for five options, you'll get five options. Think of them like working with a smart but sometimes overconfident summer intern - lots of potential, but needs guidance and verification.
One final practical limitation worth noting: LLMs can only access information they were trained on, which has a cutoff date (which they will tell you, if you ask). For anything recent, you'll need to provide that context yourself, usually by sharing relevant articles or documents.
Choosing your first (or next) LLM
The main players - ChatGPT, Gemini, and Claude - all have free versions and more capable "pro" versions that cost around 20 USD monthly. The pro versions mainly offer:
Larger "context windows" (they can handle more text at once, like entire books instead of just a few pages)
Better performance on complex tasks
Faster processing
Additional features like document upload or code generation
From my experience, while the free versions are good for getting started, the pro versions are more or less mandatory if you're serious about using LLMs for important work. The quality difference is noticeable, especially for professional tasks.
Also, the LLMs have different pros and cons:
Claude tends to be good at writing and research tasks, though you'll need to upload relevant web pages as PDFs since it can't browse the internet
ChatGPT can be enthusiastic to the point of being overly bubbly, and sometimes gives long, unfocused responses
Gemini offers solid all-around performance and works well with Google's tools, but I at least was less impressed by its problem-solving
My suggestion is to try the free versions of all three. See which one's communication style and interface feels most natural to you - and then sign up for “pro”. In the end, this is a gut decision.
How do we actually use an LLM?
The basic interaction with an LLM is called "prompting" - it's similar to having a conversation through text (or voice with some LLMs).
Let’s come back to the more strategic approach of using LLMs. Before you even open your LLM and think about the first prompts, pause first - and make sure you have done the groundwork to effectively evaluate and guide the LLM's output:
Define what success looks like. For instance, if you're working on a customer analysis, did you clarify whether you need high-level insights or detailed segments? What would a good solution look like? How detailed or specific would it be?
Reflect on your objectives. Sometimes we ask for one thing (like "summarize this report") when we really need something else, often with a clear emotional sub-level (like "I really want to impress my boss. Please help me find the three most relevant insights for my project"). Be open and honest, e.g., communicate clearly that you feel unsure about how to approach the problem, or what the problem even is. The LLM will react to that, and propose the right approach.
Think about how you'll use the output. An email draft needs different qualities than a strategic analysis over 20 pages.
Without this clarity, you might find yourself nodding along to impressive-sounding but ultimately unhelpful responses. Or even more frustrating, you feel that the output is not good enough, but don’t know what exactly the problem is, and how to improve the output.
After completing the ground work, open your LLM - and get started with your prompting workflow:
Start with a detailed initial prompt (I'll share specific guidance on this below)
The LLM responds with its initial thoughts
Evaluate the response and provide feedback based on your predefined goals
The LLM adjusts its response based on your guidance
Continue this dialogue until you achieve your desired outcome
Finally, validate and refine the results
Let's look more closely at crafting that initial prompt:
Take about five minutes to plan your prompt, or longer if needed. Investing time here will save you significantly more time later - and will substantially improve output quality
Begin with setting the context:
Define roles: Clarify your role (e.g., job, role in the business, role in the side project) and the role you want the LLM to take on (smart assistant, savvy entrepreneur, ambitious CFO, wise counselor). Describe your target audience and their needs. I've found that role descriptions have a surprisingly large impact on output quality and the style of the interaction.
Next, define your industry or area of expertise and its key characteristics. Explain where you stand today, including what has recently happened to trigger this task. Then outline your goal state - where would you like to go with your LLM answer?
Specify any constraints, e.g., time limits, budget limits. Share what you have tried already, and what didn't work.
Be explicit about what exactly you want to receive. For example, do you need a simple long-list, or do you want to start an exchange with the LLM? Consider whether you want 5 ideas or 25 - fewer ideas might be quicker to review, but a longer list could spark unexpected insights.
Include preferences about style and format - such as the tonality you would like the LLM to use (more informal, more professional), and the length and format of the expected output. For instance, do you want a list of ideas, or a fully formulated long-form article? Are you looking for a rough storyline of a PowerPoint deck, or the full wording for each slide?
Finally, present the actual task. You can take one of two approaches:
Either outline the desired end state or target outcome, then ask the LLM for suggestions on how to tackle the question. You can then fine tune the approach together and work through it step by step
Or propose your own approach. Map out the steps you want the LLM to take, and ask it to start with the first one. For example, ask the LLM to break down the problem first, then address each component separately
In practice, much of your success will actually come from developing good instincts. During a prompting session, you'll often rely on gut feelings: Does this answer really address my needs? Will it help my audience? Is it genuinely insightful, or just superficially impressive? This isn't a purely rational process - it's about developing a sense for what good output looks like, much like developing a taste for good writing or effective problem-solving.
At the same time, you can't skip the foundational work of preparation and structure. Without clear objectives and a well-thought-out approach, even the best instincts won't lead to consistently good results. You need to spend time defining your goals, understanding your audience's needs, and mapping out how you'll use the LLM's output. This structured preparation helps you ask better questions, provide more focused feedback, and ultimately get more valuable responses.
Getting started - super practically
Here are two simple exercises I've found helpful for getting comfortable with LLMs. I've chosen personal use cases deliberately - you'll have a better sense of what makes a good answer when it's about something familiar to you.
1. The "Holiday Advice" exercise: Tell the LLM about yourself - your interests, preferences, favorite activities, whether you prefer cities or nature. Ask for travel recommendations. Then refine those suggestions by adding constraints like budget or timing. It's a low-stakes way to practice the back-and-forth dialogue.
2. The "Forgotten Hobby" exercise: Think of an activity you used to enjoy but dropped. Ask the LLM to help you think through getting back into it. This helps you experience how LLMs can break down complex decisions into manageable steps.
If you know someone who could benefit from this post, please consider sharing it with them.
Also, as always, I'm interested in your thoughts: Feel free to message me. Or, if you prefer, you can share your feedback anonymously via this form.
All opinions are my own. Please be mindful of your company's rules for AI tools and use good judgment when dealing with sensitive data.