Matthieu Napoli

lol

Happy to hear!

Preparing for LaraconEU
I have now 2 floppy disk readers, I can create twice as much 💪 Come find me if you want one!

Bref 3.0 is released 🎉 – Bref Bref 3.0 is released and reaches 40 billion invocations every month.

Bref 3 is released 🎉

Since its creation, Bref has served 1.2 trillion requests and jobs!

▸ simpler & lighter runtimes
▸ PHP 8.5
▸ improved observability (logs & tracing)
▸ simpler Laravel onboarding
▸ postgres & redis built-in
and more ⤵
bref.sh/news/03-bre...

Working on Bref v3 release for tomorrow 🤞

Most emails I receive about open source are unsolicited feature requests, but not all emails 😌

AWS Lambda has a 250MB limit for applications. Reducing package size involves modifying composer.json to exclude unused services.

Improved Bref docs on a quick win: cleanup the AWS SDK before deployment
bref.sh/docs/deploy...

(btw could also be useful if you deploy your app on a server)

I don't understand why PHP got typed constants. Constants never change, their type is exactly what the constant is.

New UI to configure private networks in bref.sh/cloud

Claude Code overview - Claude Code Docs Claude Code is an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools. Available in your terminal, IDE, desktop app, and browser.

Identify rules in the CLAUDE.md files that can be turned into automated checks. Every rule we can remove is context freed up for the agent.
```

Claude Code overview - Claude Code Docs Claude Code is an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools. Available in your terminal, IDE, desktop app, and browser.

Replace Claude rules with automated checks, run this now this is so good:

```
Read all CLAUDE.md files in the project, along with the existing tool configurations (eslint, phpstan, pint, package.json scripts, git hooks, and everything else relevant).

Yeah! Especially since by moving to Prism I was able to set many "lower level" options (temperature, media resolution, etc.)
I don't regret refactoring to Prism for this use case, less abstractions, made more sense.

I used the new Laravel AI SDK and made a dumb mistake: it's really for implementing agents, not prompting models.
I had a task with a single request to send to Gemini (transcribe this image into text) and using Prism directly made much more sense in the end. TIL!

I don't know if they did "talk" to each other tbh, I haven't done anything specific. Maybe they talked about me behind my back and I wouldn't be the wiser 😛

`interview` skill `interview` skill. GitHub Gist: instantly share code, notes, and snippets.

Here it is: gist.github.com/mnapoli/3a8...

⚠️ I do use my `/interview` command and it might be important here. Link to the skill below. It makes Claude Code spend ~5 minutes with you asking questions and clarifying the task. SUPER helpful to spend less time adjusting down the road.

I think I need to adjust a bit my approach now: be less afraid of larger tasks and use agent teams for this.

I usually stop and re-scope tasks when I realize I'm going down a rabbit hole (task too big), but here I didn't and ended up implementing 2 very large tasks with an agent team with very little oversight.

Trying out the new "agent teams" in Claude Code.
I had very low expectations (I plan small tasks, "no need for a team"), but I am blown away so far.

E.g. I set GPT 5.2 at a test refactoring task (cleanup), it worked ~3 hours, had to correct course (re-plan) twice, didn't let it finish because it was absolute crap.
I asked Claude Code the same thing and it finished it without problem (~30 minutes)

I tried the Codex App but didn't have a great experience.
I tried both GPT 5.2 and GPT 5.2 Codex (high or ultrahigh), they are incredibly slow.
(that was just before the release of GPT 5.3, dunno if that's different now)

Claude keeps writing transactions wrong, I wrote this skill in case you're interested
gist.github.com/mnapoli/0e8...

(also if you spot something wrong let me know 🙏)

In Laravel apps that heavily use actions (e.g. most entities are created via actions), do you use factories at all?
E.g. `CreateUser` action does advanced stuff. Then `UserFactory` is not aligned and tests are all wrong.
Factories don't work in this case, right?

Right! I didn’t like worktrees in conductor so I dropped it

Breaking into small tasks is just an implementation detail to me (I care about the overall design & outcome), whereas when planning large features I want to read the plan. So beads was just a bit limited here in the end.

I don't think I've reached the perfect workflow yet, still iterating.

What I ended up disliking about beads is that they are mostly invisible for me (human), and even if I installed some plugin to "see" them, I wouldn't really read them, too much noise.

- GitHub issues for tracking large tasks
- I work with CC to create a detailed plan for each, and store the plan into the issue

- native CC tasks (new tool they introduced) for breaking large tasks into smaller tasks in a single session -> CC natively knows how to run sub-agents to:
- handle them in parallel when possible
- avoid cluttering the root context too much
- Not perfect but works

I was asked about Claude Code with Beads, might as well post here too.
I used Beads for a while:
- 1 terminal to create issues, or even epics for large tasks
- 1+ terminals to implement them

Now I moved to:

ok Claude Code has the cool factor but going all in on terminal UI is such a mistake IMO…

OpenAI launched a Codex app. Like Claude Code but as a macOS app. The UX looks incredible! Just watch the video!

Which OpenAI model to use instead of Opus 4.5?
openai.com/index/intro...