Against Prompt
This is an incredible time to be building in AI.
Every week there’s a new capability that makes you rethink what’s possible. I use these tools every day and I’m genuinely excited by what they unlock.
And yet something keeps nagging at me.
All of these tools, as powerful as they are, still start the same way: with you telling the machine what to do. They start with a prompt or a workflow you configured. The trigger is always human and their intelligence is always reactive.
I think there’s a different path. One where the system learns how you work instead of asking you to explain it. I think this distinction, which almost nobody is talking about, will define the next generation of AI products.
The age of prompt labor
When ChatGPT launched, it felt like magic. You could ask anything and get a useful answer. But two years in, the reality of daily AI usage looks different.
You open a tool. You connect your tools. You think about how to phrase what you need. You write a prompt. The output is 70% right, so you refine. You try again. You copy the result into the actual place it needs to go. Then you do it again for the next task.
For automation tools, it’s almost the same problem. You map out triggers and conditions. You test, it breaks when something unexpected happens. Your process changes two weeks later and the automation is already outdated.
The interface changed but we have the same cognitive load. We just moved it from “doing the work” to “instructing the machine to do the work.” We are now work coordinators.
This is fine for one-off tasks like summarizing a document, generating an image or an article. Prompting works when the task is self-contained and you can describe it fully in a few sentences.
But most real work isn’t like that. A founder’s Monday morning involves 80 unread emails, three scheduling conflicts, a follow-up they forgot to send Friday, an investor update that needs context from four different threads, and a new hire starting who needs six introductions. These aren’t separate tasks but a map of dependencies and priorities that shifts by the hour.
You can’t prompt your way through that because the work itself is too interconnected to decompose into individual instructions.
3 ways AI connects to your work
There are three fundamentally different architectures for how AI relates to your context and each offers a different kind of value.
Reactive tools: you bring the context
ChatGPT, Gemini... You input, they output. These are incredibly powerful for thinking, writing, researching, and creating. But they know very little about your actual work and about you. Most of the time, whatever context they have, you either typed it into the conversation or connected the tools yourself.
For a sales rep, this means: you can ask AI to draft a follow-up email, but you need to tell it who the prospect is, what was discussed, what the next steps are and what tone to use. The value is in the generation, but the context assembly is still on you.
Context-connected tools: the system reads your environment at time t
This is where things shift meaningfully. Tools like Anthropic’s Claude Cowork connect to your files, email, calendar, and documents, and pull context at the moment you engage them.
When you ask Cowork to draft a report, it reads your existing documents, cross-references your data, understands the structure of what you’ve already built. This is a real improvement because the AI doesn’t start from a blank state.
But the important nuance: it assembles that understanding at the point of interaction. It’s a snapshot and the trigger is still you.
For that same sales rep: Cowork can read the CRM notes, the last email thread and the meeting transcript before drafting the follow-up. But the rep still has to decide it’s time to follow up, open the tool, and ask.
Context-aware systems: the system observes over time
This is the frontier I’m most interested in.
What if the system didn’t wait for you to ask? What if it had been observing. Not recording everything, but paying attention the way a great executive assistant does. Learning your patterns. Understanding your relationships. Building a model of how you work that gets richer over weeks and months.
This is a major architectural difference from the first two approaches. A context-connected tool builds understanding at the point of interaction. A context-aware system builds understanding continuously, in the background. It knows that you always follow up with warm leads within 24 hours. It knows that your Monday mornings are for deep work and you hate being interrupted before 11am. It knows that when your co-founder emails about infrastructure, it’s usually urgent, but when they email about the offsite, it can wait.
None of this was explicitly configured. The system learned it by analysing intent across your context.
For the sales rep: the follow-up is already drafted when they open their inbox Tuesday morning. It references what was discussed, uses the right tone for that specific relationship, and is waiting for one click of approval. The CRM is already updated. The email is forwarded to the head of growth, because the system understood this deal needs their attention. The rep never asked for any of it.
What observation unlocks that snapshots can’t
I want to go deeper here because this is where I think most people building agentic AI are underestimating the opportunity.
A snapshot of your work tells you what’s there right now. Observation over time tells you something much more valuable: what matters, to whom, and why.
Take prioritization. A snapshot-based system can sort your inbox by sender importance or keyword urgency. That’s useful. But an observation-based system knows that this particular investor only emails you when something is wrong. It knows that your biggest client’s tone has shifted over the past three weeks from enthusiastic to neutral.
These aren’t things you’d think to prompt for. They’re patterns that emerge over time, and they’re the patterns that actually drive how good operators make decisions.
I studied neuroscience before building software and this maps directly to how human attention works. We don’t process every input equally. We filter constantly based on learned patterns. What’s new versus familiar. What’s urgent versus routine. What’s relevant to what we’re working on right now versus what can wait. The brain does this automatically, built from years of observation.
Most AI skips this entirely. It treats every input with equal weight because it has no history. A context-aware system starts to develop something like memory. Not perfect, not human, but useful in a way that no amount of prompting can replicate.
What this means if you’re building
If you’re building agentic AI right now, here’s the question I think matters most: is your system learning?
Do you design the system around a growing understanding of the user. Every interaction, every decision the user makes, every correction they apply becomes signal. The system gets better not because the model improved, but because the context layer got more representative of a specific user.
This changes what you prioritize as a builder:
The context layer matters more than the model layer. Models will keep getting better, faster, cheaper. Your moat is how deeply you understand the user’s world. How much signal you’ve accumulated. How accurately you can predict what they need before they ask.
Trust is the product, not the feature. A proactive system that gets it wrong is worse than no system at all. The calibration between acting and waiting, between suggesting and staying silent, is the core product challenge. You earn the right to be proactive incrementally. Start small, prove reliability, expand scope.
The system should suggest, (almost) never execute without approval. At least not yet. The path to proactive AI goes through a phase where the system does the work but waits for a human to confirm. Drafts, not sends. Suggestions, not actions. This is how trust gets built. The moment you skip this step, you lose people.
Where this is heading
There will always be a place for interactive, creative AI tools. Prompting is perfect for brainstorming, writing, thinking alongside a machine. Context-connected AI is perfect for complex projects where deeper understanding produces better output.
But a huge portion of our daily work isn’t creative or complex. It’s just constant and fragmented. And that’s the layer where prompting makes the least sense, because no one sits down and says “now I’m going to do my coordination work.” It happens in between everything else.
I keep coming back to an analogy that has nothing to do with technology. Think about what great hospitality feels like. You walk into a hotel, and someone has already anticipated what you need. Your room is ready. The temperature is right. There’s water on the nightstand. Nobody asked you to fill out a form, they observed and acted.
That’s the experience that work should feel like.
Not another dashboard, tool to learn or prompt to write. Just the quiet confidence that the things that need to happen are happening and when something truly requires your judgment, it finds you at exactly the right moment.
I think the next step for AI is care. Systems that observe your world closely enough to understand it and adapt to it so you don’t have to explain yourself every time you need help.
That’s the future I’m building toward.