I really wanted to like The Man in the High Castle, but I’m halfway through season 2, and I’m bored to death
Remote
Yannick
@osteel@phpc.social
mastodon 4.7.3Freelance Product Engineer at Yellow Raincoat Ltd. Founder of https://pixelwatcher.io. Organiser of @phpsussex@phpc.social. I post in French once in a while.
51 Followers
43 Following
45 Posts
Joined February 20, 2023
Twitter:
GitHub:
Nostr:
npub1h9hwa8fn4pljhdwan6uvhe0pmt4jnvn57wn0evhpfz4yupe9mzhqr746f9
Somehow missed the @mixin PHPDoc tag until now. On a PHP trait, it tells static analysers (PHPStan, IDEs) to type-check it as if it lived inside a subclass, so they don't flag calls to methods that only exist in the subclass as invalid.
There’s currently an uptick in people contacting me on LinkedIn offering to build my website. I’m inclined to feel offended.
📢 Bristol! I'm speaking at PHPSW this Wednesday (11 March).
I’ll be giving an extended version of my talk Escaping the Code Maze, introducing the Closed-by-Default Principle.
Are you based in Bristol or in the area? Come say hi!
@vohanj@phpc.social project managers, product managers, architects. All converging to some extent
Man, I really welcome FrankenPHP, if only so I never have to look at a PHP-FPM config again 🙏
My SaaS officially has a logo and a name! Can you guess what it does 🤪
Teaser page with newsletter form to be informed about launch and get early access https://pixelwatcher.io
And a quick tutorial on how to use Remotion and Claude to animate your own logo https://yellowraincoat.co.uk/blog/animate-your-logo-with-remotion-and-claude-code
Replying to
The lesson is that, while looking at other people's setups can be useful for inspiration, nothing beats getting your hands dirty and consolidating repeated practices (through skills, commands, hooks, etc.) as you go, until you get to *your* ideal setup.
As developers, we're wired to follow existing patterns and apply whatever the industry considers best practice. This only works partially for AI setups – tools like Claude Code warrant and facilitate extreme levels of customisation.
One reason I’m transitioning to the role of product engineer. Product engineers are accountable for outcomes, however they get there. I don’t want my value to be tied to specific inputs while AI is reshaping those inputs.
https://www.oneusefulthing.org/p/what-it-feels-like-to-work-with-mythos?hide_intro_popup=true
I understand some of the feedback around Opus 5 now. It definitely is more conservative, in an annoying way. Random directives it picks from elsewhere now seem to have as much weight as my own instructions.
Replying to
Note that I'm speaking as a solo builder here. I'm not sure how this would work for teams, but I'm guessing that it would involve some friction.
If you're looking for a task for Claude to do semi-autonomously (and that you can monitor from your phone, for instance), try `/doctor`. It does a lot more than it used to, and will probably flag and offer to fix things currently flying under your radar
Replying to
Once I was happy with it, I asked Claude to review the documents again and generate a list of potential ADRs using a template enforced through a skill. I validated each ADR one by one, amending the proposed drafts where necessary.
Those amendments led to further changes in other documents, all within the same session. Claude then returned to drafting the ADRs without ever losing context.
Pretty neat and straight-to-the point explainer of the difference between jobs, actions and services, with good examples. Only thing I would change is make the action invokable @radenadri@medium.com
I've also been using the `/grill-me` skill instead of plan mode. So far I like it – the model asks questions one by one, which feels like more work, but in a good way: it forces you to focus on one aspect at a time, instead of glossing over a verbose plan
https://github.com/mattpocock/skills/blob/main/grill-me/SKILL.md
Currently translating a presentation from English to French. It makes you realise how much more concise and efficient the English language is. French is completely messing with my slides' layout 😩
Always interesting to see how these companies use AI. I'm still not completely sure how to approach the review/verification side of things, but I totally agree on how it allows companies to scale more slowly https://replit.com/blog/self-driving-company
A great summary of the current geopolitical and macroeconomic environment, and how to navigate it from an investment perspective https://www.lynalden.com/june-2026-newsletter/
Replying to
Laravel Boost AI guidelines skill https://github.com/osteel/.claude/blob/main/skills/laravel-guidelines/SKILL.md
Anyone following the AI space has come across the term "evals", usually in the context of training models. But what do evals look like on the application side? This article covers why I ended up needing them for PixelWatcher, and how I implemented them: https://yellowraincoat.co.uk/blog/what-are-evals-anyway
Guess who’s speaking at Laravel Live UK in June 😌 https://laravellive.uk
Interesting take on why the EU regulates Big Tech the way it does, and why the US doesn’t https://thebulletin.org/premium/2026-09/three-reasons-europe-learned-to-distrust-big-tech-before-america-did/
Claude Code tip: To maximise the number of 5-hour windows I get per day, I set up a dummy routine that runs at 5:30 am. It starts my first session before I wake up, so when I begin work I have a fresh session with a couple of hours left, and can fit three sessions into the day.
I took part in the State of PHP survey by @thephpf@phpc.social and PHPStorm – join now too and help us map the herd! 🐘 https://jb.gg/stateofphp26
I finally have some time to dedicate to PixelWatcher (I’m busy with a new contract, which is good!). I’m aiming for a public launch in the next few weeks (meaning there’s still time to join the private beta 👀) https://pixelwatcher.io
Replying to
I then asked it to add some custom Laravel Boost AI guidelines via another skill, and to reference the ADR directory for scanning and progressive disclosure, giving agents additional relevant context when working on tasks.
***
I'm still very much experimenting here, but I can see how working with Claude (or your provider of choice) can lead to very powerful customised workflows.
I'm currently trying to test more LLMs for PixelWatcher. Turns out it's very hard to find reliable providers: Scaleway times out a lot, OpenRouter is even worse. I can't even create an API key on evroc. Looking into Hugging Face now, but their interface is hurting my brain
I'm finally using a task tracker for PixelWatcher (I was using a notebook until now 😅). I settled on GitHub Issues, since I'm already using GitHub anyway, and I don't feel the need for something like Linear at the moment. Claude can talk to GitHub via `gh`, which is good enough.
Fine-tuning my slides ahead of Laravel Live UK later this month 👀
🎟️ Grab your tickets: https://app.tickettailor.com/events/laravellive/2003132
caffeinate -d
Replying to
Skill to create skills https://github.com/osteel/.claude/blob/main/skills/new-skill/SKILL.md
After using Opus 4.7 for a few days, I feel the same as everyone else: the model feels less smart overall (and has a slightly more annoying personality), but it's *much* better at keeping track of multi-step workflows, which seems to confirm it's geared towards agentic work.
I thought switching back to VSCode from Cursor would also mean fewer updates... I was wrong 🙄
I'm halfway through @mattstauffer@rtsn.dev's first PragmaticAI podcast episode with Aaron Francis. Good stuff so far (appreciate the balanced views), but what immediately caught my attention is the positioning of Tighten as a company taking AI ideas and prototypes to production. Smart move!
Replying to
@ian@phpc.social so this isn’t for coding work, but as part of some application logic. I’m basically doing image comparisons, and I’m struggling to find reliable providers to try out Chinese models (I’ve exhausted most "western" models by now). But it seems to be related to image size more than anything (although the models I’m testing aren’t supposed to have any size limits)
Replying to