·4 min read
Featured

You don't need your PC terminal to get the best from AI

Everyone keeps talking about having to learn Claude Code, but that won't be the way most people interact with AI in the future

AIGeneral
Share:

Get posts like this directly in your inbox.

Subscribe on Substack

There are a lot of posts on LinkedIn and Twitter right now saying everyone should learn tools like Claude Code and start automating work from their terminal.

I don't think that's right at all.

I use Cursor and Claude Code a lot and find them very useful for coding work. But that's because a lot of coding work is happening in the Terminal anyway. Most other work isn't.

That's why I don't think it's the long-term interface most people/businesses will use.

It feels more like when websites were still mostly built by hand - if you wanted something decent online, you either learned a bit of HTML or paid someone who already knew it. That didn't mean the future of the web was everyone sitting there hand-coding websites forever. It just meant the nicer way of doing that work hadn't been invented yet.

I think agent tools are in that phase now.

The product layer is still catching up with the technical layer - and a lot of companies in the AI race are from the West Coast of America, the home of tech-based startups who live in code.

So I think the more useful question is not "should everyone learn Claude Code?"

It's: what are the useful parts of my AI system that I own?

That matters much more.

A lot of current tools look great in a demo. They have a smart interface, a name based on some Greek/Roman God or other, a few clever examples and usually a workflow builder with loads of boxes and arrows.

But if the useful bit lives inside their product, you haven't built much of an asset for your business.

What would you still want to own and if all the tools changed in a year?

  • your process
  • your review logic
  • your examples of what "good" looks like
  • your history of what has and has not worked
  • your team's judgement

That's the layer I spend most time thinking about. It's also the layer a lot of people still spend the least time thinking about in favour of the latest hot thing.

The terminal is not the point

I can imagine some people reading the last couple of newsletters and thinking, "It's fine for you to be doing all this coding with AI models, but I don't know how to do that and don't want to" - Fair enough!

I don't think most people or teams will need to.

What they will need is a clear view on a few things:

  • what knowledge in the business actually matters
  • what an agent should be allowed to do
  • where a human needs to review
  • what should be remembered for next time

Getting all of that together in a way that's useful for agents is the hard bit. Terminal-based tools is just where some people happen to be activating it early.

That is also why coding agents matter beyond software. They show the best agent pattern very clearly: good context going in, tools available when needed, human review in the right places, then another pass until the work is complete. That pattern will show up elsewhere too.

The durable layer is above the model

One reason people get distracted is that the model conversation is a much simpler and higher level one to have. Should we use GPT? Claude? Gemini? Whatever turns up next month?

Reasonable question. Not where I'd focus most of the energy.

Those models will keep improving. Pricing will move around. Different models will have different strengths and weaknesses. The "best" option for one job won't stay the best forever, and it probably won't be the best option for every job anyway.

So if your whole setup depends on picking the right model and sticking with it, that's not a good bet.

The more valuable layer of infrastructure sits above it:

  • context
  • permissions
  • review
  • memory
  • standards

If you get that layer right, swapping the model underneath becomes quick and easy, not a massive setup change.

If you don't get that layer right, changing models won't save you. You're still feeding a generic system generic inputs and hoping for a non-generic answer.

That is a big part of how I think about Praxis too. The interesting problem isn't "how do I put the latest model in front of someone?" It's "how do I make a business's actual expertise usable inside a system, while keeping what's going on behind the scenes visible?"

It was never about the tools, it was always about the information going into them.

All Posts