A lot of people use Large Language Models (LLMs) as if they were just faster search engines or copywriters. You ask a question, get an answer, copy it, and move on. After years of building AI systems, writing books, and mentoring engineers, I’ve seen that this habit often leads to a plateau. If you only use AI to get answers, you end up giving away your own critical thinking. When I started using LLMs as a thinking partner instead, my whole workflow changed.
Now, instead of letting the model do all the work, I treat it like a tireless and knowledgeable teammate. Its role is to challenge my thinking, test my technical designs, and help me spot any blind spots.
In this article, I’ll show you how I use LLMs in my daily work as an engineer, author, and problem-solver. I’ll also share the exact workflows and prompts you can try right away.
4 Practical Ways I Use LLMs as a Thinking Partner Every Day
1. Stress-Testing Architecture and Challenging Assumptions
Before I settle on an architectural design or data model, I test my assumptions using a ‘Red Team’ prompt. As engineers, we often fall in love with our first ideas, but confirmation bias can be a big problem.
I start by writing a short summary of my plan and listing any constraints. Then, I ask the model to act like a skeptical principal architect.
Here’s an example prompt:
Here is the system architecture I am proposing for our vector search service: [Insert details]. Act as a cynical Principal Systems Architect. Identify 3 points of failure under peak load, 2 operational bottlenecks I have overlooked, and challenge my database choice based on indexing overhead.
This exercise often brings up edge cases, such as re-indexing delays or tricky cache invalidation problems. Catching these early saves days of debugging later in production.
2. Exploring Trade-Offs in Code and Refactoring
I don’t usually ask LLMs to build full features from scratch. Instead, I use them to talk through my ideas and explore different design patterns. When I’m working on a complex algorithm or improving a data pipeline, I like to compare different ways to solve the problem.
I share my code with the model and ask it to review things like memory use, readability, or scalability.
Here’s an example prompt:
Review this PyTorch data-loading pipeline. Do not simply rewrite it. Walk me through the trade-offs between my current multiprocessing approach and using custom iterators with memory-mapped files. Which approach handles variable-length sequence batches better without spiking RAM?
This way, I stay in control of the decisions. I make the final call, but I understand all the trade-offs involved.
3. Dialectic Research and Counter-Argument Synthesis
When I research new AI architectures, like speculative decoding or state space models, or when I prepare a roadmap for you, it’s easy to find sources that only talk about the benefits.
To get a more balanced view, I ask the model to argue against the popular opinion.
Here’s an example prompt:
I am evaluating whether to implement a Retrieval-Augmented Generation (RAG) system with a graph database (GraphRAG) vs. a traditional hybrid vector/keyword search. Give me the strongest case AGAINST using GraphRAG in an early-stage startup environment. Focus on maintenance cost, latency overhead, and data pipeline complexity.
This method helps you see past industry hype and base your technical opinions on real-world experience.
4. Structuring Complex Decisions with Decision Matrices
When I’m choosing tools, frameworks, or business strategies, it’s easy to get overwhelmed by too many priorities. I use LLMs to help organize my messy thoughts into clear decision frameworks.
I write down all my thoughts, team constraints, and requirements, then ask the LLM to turn them into an evaluation matrix with weighted scores.
Here’s an example prompt:
I am deciding between hosting an open-weights LLM (like Llama 3) on our own VPC versus using a managed API endpoint. My constraints are: strict data privacy, a small DevOps team of 2 engineers, and unpredictable query spikes. Synthesize this into a structured decision matrix with evaluation criteria, pros, cons, and a risk mitigation strategy for each option.
Continue Learning: Master GenAI, LLMs, and AI Agents
If you want to learn more about the ideas in this article, I suggest mixing hands-on practice with structured learning.
My book, Hands-on GenAI, LLMs and AI Agents, helps you learn about Generative AI, LLMs, and AI agents by building real projects. It covers topics like LLMs, RAG, AI agents, multimodal AI, and more.
If you prefer a structured course, Generative AI with Large Language Models by DeepLearning.AI and AWS is a great choice. It explains how modern LLMs work, how they’re trained and fine-tuned, and how to build real generative AI applications.
If you want to build real-world AI skills, these resources can support the thinking workflows I shared here and help you go from understanding AI concepts to actually using them.
The Takeaway
The point of using AI in your daily engineering work isn’t to stop thinking. It’s to boost your curiosity and make your work even more thorough.
If you use LLMs passively, your skills can get weaker. But if you use them as an active thinking partner, they reflect your logic back to you, reveal hidden assumptions, and encourage you to consider other options you might miss.
I hope you enjoyed this article about how I use LLMs as a thinking partner. For more AI and machine learning tips, feel free to follow me on Instagram.
