Do things that don’t scale, and then don’t scale

MEDIUM.COM

One of the things I am most interested in is how this technology (AI for software engineering) is going to transform our industry. In this blog post Adam Derewecki shares his views, via an interesting spin on Paul Graham’s classic mantra that startups should “do things that don’t scale” - i.e. focus on shipping products, making technical shortcuts, rather than creating a scalable platform.

Thanks to AI tools like GPT-enabled coding, projects can stay intentionally small, personal, and sustainable rather than being forced to scale. Sometimes the most rewarding work is building something for a very specific audience, whether that is a private Slack community of a hundred people or a postcard-sending service for a family member. Scale does not always equal success; lasting value can come from creating tools that are satisfying and perfectly suited to a niche.

He illustrates this idea with examples of tiny, purpose-built tools such as a simple postcard mailer or a voice reminder system powered by Twilio. These projects deliver real value without demanding growth, public exposure, or heavy support infrastructure. They remain easy to maintain, safe from outside scrutiny, and enjoyable to build. In a world that constantly prioritises scaling up, Derewecki argues that there is power in staying small, focused, and quietly impactful.

AGENTS.md

AGENTS.MD

OpenAI (and friends) have created a “simple, open format for guiding coding agents”, called AGENTS.md - I’ve got mixed feelings about this initiative. In essence, it is nothing more than a text file that certain agents will add to their context. It isn’t a specification, or file format. It is just a filename. The website really doesn’t make this clear, implying it is something more sophisticated.

I’m not convinced of the value of having a separate AGENTS.md and README.md. LLMs are great at consuming content designed for humans, having separate guidance for agents and humans complicates matters.

The hierarchic support is weak, with the agent just using the AGENTS.md file from the closest folder, as a result, guidance needs to be duplicated across agent files. It would be trivial to supply an agent with multiple AGENTS.md files in priority order.

In brief AGENTS.md files are little more than a README.md file with enough hype to motivate people to create one. Are people more motivated to write documentation for machines than for humans?

📹 Senior Engineer Tries Vibe Coding

YOUTUBE.COM

While this video is a spoof, it is a good one - and very funny. Anyone who has spent a significant amount of time vibe coding or leaning hard on AI dev tools will be able to relate.

“Don’t use Javascript … npm install ?!!! What did I teach you!!”