The question you wouldn't have asked

The majority of questions people ask AI tools are questions they could have Googled. "How do I give feedback to someone who isn't meeting expectations?" "What's a good opening for a cold email?" "What are the main risks of starting a subscription business?" These are fine questions. They get decent answers. But they're not the real questions — the ones about your actual colleague, your particular business, your specific history with someone. Those stay unasked, because you already know the AI won't have what it needs to answer them.

Rich personal context changes this. Not just answer quality. Question quality.

The grammar of the generic question

There's a useful pattern in how people prompt AI without context: the request-only prompt, where someone asks a question without providing any surrounding information. Nielsen Norman Group documented these as resembling search-engine queries — efficient at getting generic answers, requiring significant back-and-forth before they become useful. People know, instinctively, that specific questions require the other party to have context. So they preemptively adjust, stripping the specificity out before they even ask.

This isn't laziness. It's a rational response to a perceived limitation. When you believe the AI can't answer a contextual question, you don't ask one. The constraint lives in your model of what's worth asking, not in the AI's capabilities.

James Gibson, the psychologist who introduced the concept of affordances in 1979, described them as what an environment "offers the animal" — the action possibilities it makes available depending on the abilities of whoever is engaging with it. Donald Norman extended the concept to designed interfaces in The Design of Everyday Things (1988). The logic holds for AI tools too. A tool with no context about you affords generic questions. A tool with structured knowledge about your work, your relationships, and your history affords specific ones. Different tools invite different interactions. Most AI tools, by default, invite search-engine behavior.

The question I hadn't thought to ask

I kept notes on a project that had been going sideways — not task lists, but the actual arc: what we'd tried, where the disagreements had been, what I thought the other people involved cared about. Structured enough that an AI could retrieve and reason about it.

The first time I asked about a meeting I was walking into, something changed. Instead of "how do I approach a conversation where there's been tension around scope," I found myself asking: given what I know about where this friction came from, is there a way to reframe the resourcing question that doesn't feel like a retreat?

That's a different question. Closer to the actual problem. More likely to produce something I can act on.

The bottleneck hadn't been the AI's reasoning. It was the question itself.

What narrowed questions cost

There's a version of this that runs the other way, and it's worth naming. When tools can only answer generic questions, your thinking gradually adapts to match. You learn to frame problems at the level the AI can respond to. The specific, contextual parts get trimmed away — not because they're unimportant, but because you've learned they won't get traction. Over time you stop noticing the trimming.

The hidden cost of context-blind tools isn't bad answers. It's narrowed questions. When you ask "how do I manage a difficult stakeholder," you're guaranteed an answer that's right for someone else's stakeholder. Yours has a history. Yours has already pushed back on the two most obvious approaches. Yours has one specific thing they care about that changes which framing will land.

The question that accounts for all that is harder to ask. It requires context you'd have to assemble yourself, in the moment, before you could even formulate it. Which points to the real design problem on the other side of the equation: the value of a knowledge base isn't in how much it stores, but in how much of it is structured enough to be available at exactly the right moment.

The test

The question I've started applying to my own notes isn't "did I capture this?" It's "does this let me ask the question I'll actually have?" A person record with a name and an email address doesn't afford any question worth asking. A person record with relationship context, a few notes on recent history, and something about what they tend to care about affords several questions you might not have thought to frame.

The strange thing is that better context makes useful questions feel more personal, not more general. Most software pushes the other direction — toward abstraction, toward the reusable and templateable. But the questions most worth asking an AI in practice are exactly the ones that couldn't come from a template. They only become available because someone did the work of building the context they require.

That's a different value proposition than "better answers." It's a claim about what questions exist for you at all. And the questions available to you shape the thinking you do.

---

Asgeir Albretsen is the founder of Harbor.