Artificial Intelligence, as a publicly used tool, is still in its infancy. It’s a big-ass baby, but still very young.
So, when people say it doesn’t work, they can get all sorts of advice. Last summer, the big thing was prompt engineering — and it’s still important, now evolving into what’s called context engineering. The way you ask for something and the information you provide to support that request drive the results.
Providing context in real time when making requests is important. But building a library of context over time is where the real gains happen. Here is a solid breakdown from Anthropic on how to get better results through context.
AI Is Not an “It”
As the tools have evolved, and our use of them matures, we’re seeing real differences in results — and not always for the reasons people think.
First, we need to stop treating AI like it’s one thing. AI is not an “it”; AI is a “they.” There is no monolithic AI, and keeping this in mind helps when you’re trying to get better results.
Even among the popular LLMs, there are meaningful variations — and they’re intentional. Within the Anthropic family, you have Opus, Sonnet, Haiku, and Claude Code. With OpenAI, there’s GPT-4o, o1, o3, 4o mini, and Codex, among others. Then there’s Grok, Gemini, Copilot, Llama, and more.
These aren’t just different flavors. They were designed for different purposes.
Choosing the Right LLM
The solution you’re trying to achieve should drive which LLMs you use — and you may need more than one depending on your objectives.
The obvious examples are the coding-focused AIs. It’s right in the name. Their intended use case is clear, and they’re damned good (we use them a lot). But matching other tasks to the right model takes more thought.
Going beyond basic consumer use cases — “where can I get good sushi near me” — you start to think about what you actually need AI to do and how much reasoning is involved.
The Workhorse Tasks
Some jobs are high-volume and low-intensity. Think about things like interpreting a question in a defined topic area and responding quickly with the right answer (AI chatbots, AI email responders), repeatedly evaluating and organizing data sets (monthly transactions, website trends), or handling other routine tasks that follow a predictable pattern.
These are well served by lighter, faster models like Claude Haiku, Claude Sonnet, or GPT-4o mini. They’re built for speed and efficiency.
The Thinking Tasks
Higher-level activity — research, analysis, reasoning, strategy — warrants the more capable models like Claude Opus or GPT-4o. These models were trained to take more time, examine information, and deliver a considered response.
They’re also surprisingly good at consulting.
The Underrated Use Case: AI as a Consultant
The consulting aspect of AI is seriously underrated. At the very least, these models make a great sounding board. But it goes well beyond that.
With a well-constructed prompt and solid context, you can engage an AI in a lengthy, productive exchange of problem-solving. We’ve done this with client strategy, campaign planning, and technical troubleshooting. The depth of the conversation scales with how much context you provide.
And here’s the thing — that problem-solving conversation can include, “Why are your answers so crappy for me?”
If It’s Not Working, Don’t Give Up
If your experiments with AI aren’t panning out, don’t walk away from it. Instead, step back and evaluate. Look at your use case. Align it with the right models — and try more than one. Experiment with different approaches.
If it’s still not getting you what you need, literally ask it: “How can I help you help me solve this?” You’d be surprised how useful that conversation can be.
The past year has shown us that these models evolve fast. What didn’t work three months ago might work today. The models’ ability to handle large amounts of context, retrieve information accurately within that context, and connect ideas has increased dramatically in just the past couple of months.
The point is, if you can’t get what you want from AI today, don’t stop. It likely won’t be long before you can.