I’ve been wanting to redesign the interior of my campervan. My current setup is a simple timber rock’n’roll bed with a mattress from an IKEA Lycksele Håvet sofa bed. It’s very comfortable, but it takes up the entire cargo area, and switching between bed and seat mode means moving all my stuff around.

I’m in the luxury position of being able to work remotely, so I’ve been dreaming up a layout that makes better use of the space: a narrower bed that’s easier to convert, and more storage in a fixed cabinet unit.

Even though I haven’t written a single line of code by hand, I have produced lots of software and finished several projects this year. AI-first development has fundamentally changed what it means to be a software engineer, and I feel we’re just at the beginning of this revolution. With every new state-of-the-art model release I’m getting closer to the point where my markdown documentation with instructions for Claude Code is my source code, and the actual code is a build artefact, just like compiled executables are a build artefact of higher-level languages.

In a conversation I had recently, someone suggested that our human thinking at some level is the same as LLM thinking. I’ve thought about that a lot (over the past decades, given my degree in AI from many years ago), and it’s a really interesting question: What actually separates human cognition from LLMs? Is there a fundamental difference between human and machine intelligence?

I think that there are, at least currently, a number of big differences between the way we think and the way LLMs think.

I regularly spend time optimizing my CLAUDE.md, and I often ask claude in an interactive brainstorm session to evaluate the effectiveness of certain prompting techniques. I have iterated over this for a few hours this evening and I thought I’d share my notes:

• All caps and **bold** have a small effect, but document structure and placement (at the top = more important) carry much more weight. Structural markers such as <CRITICAL>...</CRITICAL> are also effective.
• emojis in the prompts have a small effect, but the actual words they prefix are much more important
• Repeating the most important rules (e.g. at the top and bottom of CLAUDE.md) make it more likely it will follow those instructions.
• Tell claude that if it makes a claim, it has to show its output, e.g. “Mandatory: run all tests. Show test results in output.”
• An effective way to define behavioural constraints is:
  • BAD BEHAVIOUR: [describe a bad thought pattern]
  • REASON IT’S BAD: [describe why it is bad]
  • DETECTION: [how does claude know it’s falling into this pattern]

As an example, I’ve got this at the start of my CLAUDE.md, before any project-specific instructions:

Here’s a couple of git aliases I use all the time. It gives you a quick and clear overview of your branches and commits. It can even watch for changes in realtime and show you any changes.

A better git log

Add this to your .gitconfig:

1
2

[alias]
    tree = log --graph --decorate --all --format='%C(auto)%h%d %Creset%s %C(dim)(%an, %ad)%Creset' --date=relative

Run git tree and you get a compact view of your commit history with branches, colored hashes, commit messages, and authors. The --all flag shows all branches, not just the current one.

I have hardly written any code manually in the last year, yet I’ve built heaps of smaller and larger projects, using AI to learn new things, write boilerplate code, resolve tricky bugs, and rapidly iterate on ideas.

I have journeyed from AI autocomplete to vibe coding to prompt engineering to increasingly effective agentic spec-driven development workflows. For a while, I’ve been using and developing my trusty old prompts from my cursor-rules repo to help me keep the AI on track.

I do all my development with Claude Code and the spec-driven workflow du jour. Claude is usually great at letting me quickly explore new things and rapidly iterate on ideas.

If you’ve used Claude Code, you will be as tired as I am to have to give Claude permission to do basically anything. Can I edit this file? Yes, and don’t ask me again for similar commands. Can I run this script? Yes, and don’t ask me again for similar commands, etcetera.