Raphaël Simon

The more I use CLI LLM tools like Claude Code and Opencode, the more I realise that : - the skills specification is one of the best ideas I’ve seen and used until now - it does exactly what packages do with my favourite languages : factorise some cool stuff into reusable chunks that I can call with very few words - it’s still not mature and it should definitely evolve to behave like packages

I’ve been experimenting with ideas that revolve around this skills as packages approximation and I like the current results enough that I want to talk about it, if only to read it back again in a few months and see how bad this take is. My current take on this is an agent kit manager (akm)

LLM Skills, as in the open specification for a reusable prompt format, are CLEVER. Everyone was playing around with prompt engineering and reusing bits that worked in larger and larger prompts, copy and pasting things, cross referencing structural documents. Skills deliver on all fronts : progressive disclosure lets us save some context for the good stuff before landing in the dumb zone, self-activation means we don’t have to waste turns or tokens naming the invocation, and the folder structure means we can have even more progressive disclosure, with files read only in certain situations, scripts used as part of the skill … ...

Since the end of December 2025, I’ve spent 10-12 hours a day writing software with LLM coding agents and CLI-tools.

I’ve kept track of the experience from day one of trying out Opus 4.5, across notes on my phone. This post is about what it actually feels like: going from an e-bike to a race car.

In the first 3 quarters of 2025, LLMs were as useful to me as they had been since OpenAI unleashed ChatGPT on all of us. I felt like nobody had cracked a sustainable workflow. Lots of copy and paste from the web chat UI. Cool demos and some attempts at more autonomy but lots of horror stories (bye bye production DB). Models hallucinated just as much as they used to, too much for production trust. The output wasn’t far enough ahead of what a competent human could produce, in most instances.

Two things changed, almost at once.

1. Models got meaningfully better. Not incrementally. With the release of Anthropic’s Claude Opus 4.5, suddenly we got better one-shot successes. Better coherence across long plans. Fewer moments where the model drifts away from what you asked.

2. CLI agents like Claude Code (or Copilot CLI, Mistral Vibe, Opencode…) gave models access to the machine, in the right way. grep, find, git, CI/CD: bash is all you need apparently! The model can explore code without wasting tokens. It can participate in versioning its own work, and documenting code. It can run tests, read failures, try again. It could do all of this in parallel across branches, or git-worktrees.

Welcome to my blog! This is where I’ll share my ramblings, on anything ranging from medicine and public health to software engineering, data science, and the interesting problems that emerge at their intersections.

I’ve been trying to run a blog since I was 11, that’s more than 20 years ago. What happened now ? I don’t know. I guess I got old enough to realise how important writing is ? Or maybe I’m just in a phase that will last two weeks.

Who knows ?