AI

My Claude Code setup

9 commands, 5 agents, 10 skills, 4 hooks and 8 mods.

A warning: this article is fairly technical. If you are not into AI, software development or code, if GitHub, Claude Code and deploying applications are all Greek to you1, the rest will very likely go a little over your head. That is not a big deal, and I will still do my best to explain a few concepts.

First of all, “Claude Code” is the name of a tool — and what a wonderful tool!, and I realise that this alone is not obvious to all my readers. So the footnote here2 holds a few definitions that may help with the rest of the article.

All introductory considerations exhausted, on to the heart of the matter. Quite soon after GPTs, LLMs, AIs or whatever you want to call them appeared, the idea emerged that these tools are rather good at writing computer code. Which makes sense: they are trained on public data, among other things, and there is an enormous amount of public code. In the same way that comparing a French text with its Korean equivalent can train an AI to translate, AIs could be trained to translate natural language into computer code (I am simplifying a little, I admit).

Which, as far as I am concerned, is excellent news. I have plenty of ideas for applications, I quite like writing code, but I am terribly short of time for it. So if I can just tell an AI that I have such-and-such an idea, and it writes the code itself, that suits me.

That is where “vibe coding” comes in (in French I keep using the anglicism, until the Québécois sort the problem out), along with all the tools launched on that promise: letting anyone produce an application without writing a single line of code. I recounted my experiences a little on LinkedIn, not always very conclusive ones, and then one day Claude Opus 4.6 arrived.

To be completely honest, the idea of subscribing to a tool had been on my mind for a while, I have always slightly preferred Claude to ChatGPT, and as it happens I took the plunge when that famous model came out.

In a word, I am extremely impressed by Claude Code. I used it to completely rework every project I had vibe coded with Bolt, Lovable and Replit. Claude Code greatly improved not only the code (a simpler technical stack, optimisations, security), but also helped me sharpen my ideas.

The tool lets you start with a brainstorming phase (what a shame that the infinitely more poetic French word, remue-méninges, never caught on in everyday language!) that gives you an honest opinion of the project and its quality, and sets it off in the right direction. Then Claude builds plans in discussion with the user, to make sure it has understood. And finally it gets on with the execution, calling on several agents if needed.

And in the end, I arrived at something I want to share: my Claude Code configuration. It is not the configuration of a real developer, but it is the one that lets me run my 15-plus applications very easily, with enough quality and consistency, and without falling into the usual traps of AI-written code (notably piling up technical debt, endless cycles of fixes, or purple-tinted universes).

In our day-to-day work, these tools can genuinely save an enormous amount of time. I am not a developer, but in a few hours I have perfectly valid products (for personal projects). There is potential to change the way our “Business Analysts” work with the business to gather requirements, and with developers to get them built, which I want to explore in my team.

My intuition is that we are not very far from a time when an application for a simple need will be a kind of commodity. Example: I would like an application to manage my budget according to such-and-such principles, please build it, and 15 minutes later I have something that works. The limit today is probably deployment. Even if Claude Code already handles it very well, I think you still need some notions of architecture to get a good result.

In any case, this tool supports a lot of experimentation I wanted to do, and has let me go further than I imagined. I no longer make only websites (like Untilt.app or My-tsundoku.app), I have also built a Mac application for my novel project, and reverse-engineered a proprietary file format (to inject it into an application).

Claude without a limp

Rereading these lines a few months later, I noticed that AI is not the only thing moving fast: so is my writing!

Since other pieces follow my progress as a writer, let us stay focused on AI, at the intersection of three observations:

  • the filter bubble I am in on the subject;
  • the speed at which new things come out in this field;
  • the undeniable tendency to embellish reality when it comes to communicating, especially on social media.

Sitting at the centre of this Venn diagram, I see a great many posts of every kind about artificial intelligence, from the influencer explaining how he replaced his entire staff with 5 agents, to Anthropic explaining, a little more guilelessly, how it uses its own tools. I have ended up making a habit of something I think will be useful to anyone who uses AI: asking AI how to use AI. Even for ordinary conversational use, which I think hardly anyone escapes these days3, a first prompt saying in substance

write a PROMPT to do this

improves the results noticeably4.

What I have been practising these past weeks is the advanced version of that: discussing with Claude various tools presented as revolutionary, and the methods Anthropic recommends, so as to bring only the essentials into my own configuration, what is useful given my goals. A configuration that also improves through my own findings, often through mistakes: badly written code, a security problem, going over certain limits, are so many chances to improve things together with the AI. And one that must also adapt to how the models evolve and to what they make possible: what was true in January is not necessarily true in October. Older models needed certain precise instructions that can be counterproductive with the newer ones.

Finally, and here we reach the depths of technical considerations, I have refocused my use on a terminal in macOS5, no more Cowork at all, and very little of the chat, which now only answers the questions too pointed for a search engine to solve. In particular, I made use of the extensive customisation options to create a few mods I am rather proud of: one sums up what the AI has learned about me over time, another shows the connections between ideas, yet another helps me manage usage limits, context or the deployment of agents visually.

For those who use AI little, I think the method is useful: rather than simply taking everything offered as it comes, a first round of assessment and adaptation should help get a more relevant result. Keep in mind that the models are generative: producing an answer is their reason for being, so the quality of the question and of the approach remain essential to the quality of what is generated.

And more generally, this approach is an avatar of my learning by experience and in public, which is all the more valuable for being confronting6: any feedback, general or pointed, that you can give me will help!

This piece tightens three issues of my newsletter, in French: those of February and April, and a third one forthcoming.

Footnotes

  1. Joke time: — I speak every language in the world except Greek. Ask me to translate a word into any language and I will. — All right, how do you say watermelon in Urdu? — Ah, you know, Urdu is all Greek to me. ↩

    • GPT: Generative Pre-trained Transformer, as in ChatGPT, which was the first to let people chat with a GPT, the start of the AI revolution.
    • LLM: Large Language Model, to put it simply, what sits behind a GPT.
    • “Vibe coding”: programming (writing code, building an application) using an LLM.
    • Claude: the equivalent of ChatGPT. OpenAI is the company behind ChatGPT, Anthropic is the company behind Claude. Both are AIs in the sense the general public means it today.
    ↩
  2. A thought that may well be down to the filter bubble I mentioned earlier. ↩

  3. This is an observation based on my own experience, which I generalise without shame. ↩

  4. Should there be a fan of the various terminal emulators among you: I use Ghostty, on Claude’s recommendation as it happens, which kindly pointed out to me that iTerm2 was a little dated. ↩

  5. In the French original, « confrontante », which I invite you to read as a Belgicism rather than an anglicism. ↩

Elenchus read this story

Reasoning weak

Le texte présente une expérience personnelle sur l'utilisation de l'IA pour le développement logiciel, mais son raisonnement repose sur des généralisations non justifiées et des affirmations non étayées.

What it flags — Hasty generalization Ce qui est assez logique : ils sont entraînés notamment avec des données publiques, et des lignes de code publiques, il y en a énormément. L'auteur affirme que la capacité des LLM à écrire du code est « assez logique » parce qu'ils sont entraînés sur des données publiques, sans expliquer pourquoi cette corrélation implique nécessairement une compétence en génération de code. Cela revient à supposer que la quantité de données d'entraînement détermine directement la qualité du résultat, ce qui n'est pas démontré.

My newsletter

In French: the books I read, the exhibitions I see, the AI I practise, the biases that catch me — and right now, a writer’s apprenticeship in public.

All issues