Issue 744

27^th February 2026

Written by Dave Verwer

Comment

I wrote about ambiguity in agent prompts last week, and it feels like there’s a related discussion going on at the moment in the wider tech community that I want to dip into today.

There are so many stories of people (usually managers) inside businesses launching Claude Code, and shipping a new system internally that replaces some SaaS product that their company uses. Apart from the fact that these stories are usually exaggerated by people for the LinkedIn algorithm, I also can’t imagine any of these replacement tools last very long before everyone goes back to the SaaS that was “replaced”.

Agents are undoubtedly getting more powerful, but building software has always been much more about the decisions we make about what software does and how people interact with it than the code that we write.

Even if we take these stories at face value, no software written in a week (with an agent or without) can compete with a mature SaaS service. Yes, it might help with a simple task, but does it do everything that everyone in that company used the SaaS product for? Of course it doesn’t. What I talked about last week in terms of detailed thought going into the prompt isn’t just for code. It’s much more about thinking through edge cases, user interfaces, features, and everything else that goes into making good software.

It was Gus Mueller’s latest article that prompted me to write about this topic again this week:

But at the same time I’m not worried about being replaced by AI, or by quick free apps that have been built by AI. And in some ways I’m more hopeful than ever.

For almost 20 years now, I’ve been feeling the pressure from competing image editing apps and the potential of everything falling apart and the utter doom of my chosen profession and company. These feelings are not new to me.

So I kept on making the software that I wanted to build, that I wanted to exist in this world. And though some months or years are rougher than others, people still were willing to pay for what I’m making. And that’s kept me going, and more importantly - employed.

Just for fun, let’s fast forward ten years to a world where AI-based tools take over the whole process of building software. They can gather requirements, specify solutions, write code, read and respond to customer support and feed bug reports and feature requests back into the specification process. Even if that all comes true, we still need people who care:

Maybe I’ll even have fewer competitors in the long run, or at least not as many new competitors. Because at some point it’s not about how good a programmer you are (and I’ve always been a middle-tier programmer), it’s about discipline and vision.

Without the people with “discipline and vision” in every area of software development, including coding, we’ll just end up with an average of everything, and that’s not good enough.

– Dave Verwer

Sponsored Link

Mobile testing is notoriously hard to scale: real devices vary, UIs change, and scripts become maintenance. But what happens when AI agents operate at the visual and intent layer instead? Join us to explore what an AI-native testing strategy looks like in practice and how it applies to modern mobile and web applications. Register for the free webinar today.

News

If you have any interest in embedded Swift, this announcement from Joannis Orlandos should be very exciting:

Anyone who’s worked with embedded Linux boards like NVIDIA Jetson knows the pain: hours of setup, flashing SD cards, hunting for IP addresses, configuring SSH, installing toolchains. We recently ran an experiment with an MIT robotics student who spent 3.5 hours on the traditional Jetson setup and still couldn’t get to building apps. With WendyOS, she went from brew install to running code on the device in 3 minutes.

I’ve been meaning to include this for a few weeks. I’m not sure how it got skipped. At least I’m fixing it now, though!

Tools

I really liked this app from Sayuj Suresh for connecting back to your Mac via SSH from your phone. The website emphasises that it can be used to keep Claude Code or a similar agent fed with instructions, but in reality it can be used for whatever you want to do via a terminal to your Mac. The initial setup is incredibly smooth, and once the connection is established, none of your terminal output goes through the service. 👍

For full disclosure, Sayuj sent me a license for the software.

Code

Anton Gubarenko:

If the previous session about “Coding Agents in Xcode” lasted 16 minutes and had only 3 questions, this one was completely different. It ran for more than 3 hours (!) and included over 100 questions.

This looks like it was a fantastic session, and thanks so much to Anton for writing it up so diligently.

Artem Novichkov:

Token usage tracking helps you make informed decisions about prompt design — whether to simplify instructions, reduce tool definitions, or split conversations into multiple sessions.

I’ve not seen too many people writing about the Foundation model recently, but these new APIs in 26.4 are really going to help when figuring out what prompts are going to give good results.

Peter Yaacoub:

This article started as a comparison. It ended as a reminder: persistence is not about choosing the best tool, it’s about choosing the right trade-offs.

My biggest issue between these two technologies is the inconsistency with spaces in their names! 😂 However, that’s not all you should be worrying about, and depending on dataset size, you might want to choose one over the other.

Business and Marketing

For me, taking new screenshots for the App Store was always the worst part of the release process. For Amy Worrall, who automated the whole process, it’s a breeze! I’ve linked to several workflows like this for iOS, but don’t think I have linked to one for macOS before!

Jobs

Senior Software Engineer - Mobile @ Hatch – Join a team helping 5M+ people sleep better with smart devices in 1 in 3 nurseries. You’ll own features end-to-end on real IoT products people love, backed by world-class investors, a Great Place to Work culture, and exceptional benefits including full health coverage and meaningful equity. – Remote (within US timezones)

iOS Engineer, All Levels @ Ramp – Build Swift and SwiftUI features that power real-time spend, travel, approvals, and AI finance workflows for 55,000+ businesses on a high-velocity, customer-obsessed team. – Remote (within US timezones) with some on-site work (Canada or United States)

And finally...

All this because I’m putting off packing for my trip tomorrow.