Prompting is dead. Building systems isn’t.

That headline is intentionally provocative.

Prompting still matters. A lot.

But after spending the last week in San Francisco talking to founders, engineers, and AI-native teams, one thing became very clear to me:

Prompting is no longer the moat.

The real advantage now comes from building systems around AI.

If you want a glimpse into what AI trends are coming over the next 6–12 months, I’d strongly recommend flying to the Bay Area and attending as many meetups, conferences, and builder events as possible.

That’s exactly how I spent the last seven days.

And seven more will follow.

My reality check in SF

Last year, everything here was about prompt engineering.

People talked about:

better prompts
AI wrappers
productivity boosts
copilots for knowledge work

The dominant question was:

How can AI help humans work faster?

This year the conversation feels very different.

The question now is: How can humans build systems where AI works on its own?

That shift is massive.

I started thinking about it in three lanes.

Right lane: Prompting

This is where most people still operate.

They open ChatGPT, Claude, Codex, or whatever model they use and hand off task after task.

Write this.
Debug that.
Analyze this.
Summarize that.
That still works.

It’s useful. It increases output.

But if that’s your entire workflow, you’re already in the slow lane.

Middle lane: Context systems

This is where things get more interesting.

People in this lane understand that prompts alone are not enough.

They start externalizing:

domain knowledge
preferences
decision frameworks
taste

They store everything in structured docs, markdown files, knowledge bases, databases, or internal systems.

Their big realization is simple:

AI performs much better when context is portable.

The better your context, the better your output.

Left lane: Agent orchestration

This is where things get wild.

These people don’t just use AI.

They build systems of agents.

Incoming email? → agent triggered
Bug report? → agent triggered
Slack message? → agent triggered
New task created? → agent triggered

One agent starts working. Another agent verifies. A third agent critiques. A fourth extracts learnings and improves the system.

This isn’t theoretical.

Teams here are already shipping like this every single day.

And that was probably my biggest reality check in SF.

The left lane isn’t some futuristic holy grail.

It’s already happening in large scale enterprises here.

The new builder stack

What surprised me most is how blurry traditional roles are becoming.

The line between:

Product
Design
Engineering

…is getting weaker.

PMs and designers are shipping features themselves.

Developers are not just coding anymore.

They are doing research, stakeholder interviews, design decisions, architecture, and implementation.

The separation between disciplines matters less.

What matters more now:

problem understanding
product taste
design taste
systems thinking
execution speed

The question is no longer:

Are you a PM, designer, or engineer?

The question is:

How much complexity can you handle while maintaining good taste?

So why is prompting dead?

Prompting is not literally dead.

A good prompt still matters. A lot.

But prompting alone is becoming table stakes.

Everyone has access to frontier models.

Everyone can prompt.

The real differentiator is no longer prompting.

It’s system design.

I personally think of models like GPT 5.5 or Opus as highly capable but still raw intelligence.

Like an extremely smart graduate who has no context.

With prompts, you give direction.

With context, you give knowledge.

With systems, you give leverage.

That’s the real shift.

Building systems

Building systems means going far beyond prompting.

It means designing an environment where AI can operate end-to-end.

A real AI-native system needs five components:

Knowledge: All your domain understanding in structured form.
Taste: Your standards, preferences, and decision criteria.
Prompts: Reusable instructions for specialized work.
Tools: Access to APIs, software, workflows, and external systems.
Orchestration: Triggers, coordination, verification, and feedback loops.

That’s the level builders in the Valley are moving toward.

And yes, this is still early.

Everyone is still figuring it out.

But one thing feels obvious to me:

This is the work we need to do now if we want to stay relevant over the next 2–5 years.

Otherwise, the future role for many knowledge workers might look like this:

Babysitting agents.

Approving outputs.

Being the final human responsible for decisions.

Where I want to improve.

I’m not in the left lane yet. But I’m getting closer.

So far I’ve built a system where:

most of my knowledge is accessible via markdown
I use Notion as an orchestration layer
multiple agents can work on task boards
agents can learn from mistakes over time

Still early. Still messy.

But the direction feels right.

Another trend: startup focus has shifted

I also noticed something interesting in the startup ecosystem here.

Last year’s AI startup wave was mostly about accelerating human work.

This year, a huge part of the startup scene is focused on infrastructure for agents.

Questions like:

How can agents make payments?
How can agents provision compute on demand?
How can enterprises manage long-running context?
How do we reduce infrastructure cost for persistent agents?

This is a very different market.

The bottleneck is no longer model intelligence alone.

The bottleneck is infrastructure around agents.

That’s where a lot of value will be created.

What’s next for me

This trip gave me already a lot of clarity.

I identified three areas I need to focus on next.

1. Agent Native

This means giving my agents access to everything I can do.

If permission alone isn’t enough, I need to teach them how I work.

Every workflow I repeat is a candidate for automation.

2. Harness

Harness is my system prompt layer.

The goal is to build the best possible set of prompts, instructions, and frameworks for every kind of work I do.

That creates something valuable:

Model independence.

As base intelligence across models converges, the real value will sit in your harness.

3. Claw

Claw is what happens when everything comes together.

It gives agents a heartbeat.

They know when:

a Slack message arrives
an email comes in
a bug is created
a task changes

They act. They verify. They improve.

And they only ping me when I’m actually needed.

That’s the endgame.

A 24/7 agent system operating in the background.

Final thoughts

This week in San Francisco gave me direction.

Not just on where AI is going.

But on where I need to go as a builder.

Prompting got us here.

Systems will define what comes next.

And I have a strong feeling that the next major advantage won’t come from having access to better models.

It will come from building better systems around them.