From prompts to harnesses to context
At the start of last year, everyone was talking about prompt engineering. Then the conversation moved to harnesses, and now it’s context engineering. Most of us already understand why each one matters:
- Prompt engineering is giving the model the right level of direction, through clearer instructions.
- Harnesses tie the prompt to the work. They make sure the right context shows up at the right time and the model keeps heading toward the outcome you want.
- Context engineering is giving the model the right level of background information.
Each layer made outputs better, but none of them closes the gap we’re now seeing, and that gap is the reason we started this newsletter.
The missing piece: nobody’s asking why
The quality of a model’s answers grows with the context it has and the guidance it gets, which is also how people grow.
When we’re young, we ask why constantly, and every answer pulls out another detail about why a decision was made or why something works the way it does. Over time, those answers turn into judgment.
LLMs don’t ask why; they answer with a confidence that makes it look like they know. They often know what to do, but they rarely know why.
Hold the avocado
Say I order a salad and ask them to leave out the avocado. An LLM hears that and does one of two things: it holds the avocado this time, or it remembers, “this person doesn’t want avocado.”
The second sounds smarter, but it might be wrong: maybe I love guacamole and just don’t want raw avocado in my salad.
Now picture my kid making me that salad. When I say, “Wait, no avocado,” my kid asks, “Why not?”
So I explain that raw avocado made me throw up when I was little. I don’t like it raw, but I’m fine with it cooked into something.
That one why opened up context that lived only in my head until someone asked. Now my kid knows the rule and the exception.
Corporate America runs on the what
Most companies have the same gap. The words and the SOPs are written down, but the why behind them almost never is.
That matters, because the why is what tells you when a rule bends, and that’s as true for agents as it is for people.
Introducing curiosity engineering
So here’s the idea we want to put a name on: curiosity engineering. It means building the habit of asking why into how we work with LLMs.
Every why you capture becomes context the model can reuse. The next time your agent faces a similar task, it understands the steps it’s following and knows when they don’t apply.
Why the why is so hard to get
The why lives in people’s heads, and people rarely volunteer it. Most of us are ready to dictate what has to happen, but few of us stop to explain why.
A few people, the ones with a teacher’s DNA, will walk you through the reasoning behind every step. The rest of us are moving a million miles an hour: A happens, then B, then C, then D. We don’t pause to say why they happen in that order, or why A happens at all.
What we’re building: a why over your shoulder
We’re introducing a process that fits naturally into someone’s existing workflow and asks why at the moments that count:
- When someone is evaluating options
- When someone answers a question differently than you’d expect
- When a decision doesn’t match the written process
Each answer becomes curiosity context that, once captured, makes every future output better.
The difference between good and mediocre
Curiosity context is the difference between a good output and a mediocre one, because it gives your agents the why on top of the what. We see it as the next step for development, engineering, product and business teams who want to take their agents further.
Now imagine flipping the relationship: instead of you constantly steering your agent, your agent guides you and makes sure it gets what it needs to succeed. That’s what we’re building.
If you’re curious, give us a holler.
