Philipp Reiner

So @NotionHQ injects ads into their MCP and nudges my claude to recommend notion for a task I did not even intend. I think this is bad.

Once the final full clickdummies are done, I let it breakdown into an atomic design pattern. From here on, we can rebuild the ui components and design tokens. This ensures a super clean implementation

Prompt it to also define principles, rules and do/dont patterns with examples. These are turned into design rules and linter logic later.

We iterate on small html dummies by transcribing feedback. Its important to start with components and not full pages to direct the attention of opus.

Then I let it spawn agents to explore direction. I ensure it is not reading e.g. existing design.md or other reference which limits.

I paste in around 20-30 inspiration screenshots to analyse and give it a transcribed braindump of the mindset I want to achieve. "McKinsey meets CAD software".

Opus-5.5 with the right instructions can turn the generic @shadcn look into a highly polished experience and clear code. Her e is how

Why is gpt-6.1-sol using claude cli to call sub-agents defined in the .codex/agents folder? Even Codex is now using Claude...

Did I just fight for 2 hours with AWS IAM this morning? Maybe

This opus 5.5 video thing really works amazing. First try, around 30 minutes with opus-5.5-medium and a transcribed long prompt.

Optimized my software "kitchen", got great customer feedback to act on, new models to play with and shipped 24 PRs today without issues. Thanks @poteto for the inspiration

these things give me a better feedback if a bug has been fixed than half of the meat engineers i worked with in the past (it also extended the E2E flow to check for this)

my little fable autist today again spent 45 minutes writing 120 lines of spec after.. writing a spec.

Last week the core loop of our new antasar.com worked. We can now detect project slop aka drift in scope that often kills large enterprise IT projects.

I need to stop scrolling and post more (inspired by @thekitze). So timeline disabled, only posts are possible.

I call my GPT-5.6 my little autistic friend. It needs precise instructions but seeing it run for hours straight, producing high quality code is amazing.

There is not a single week without a "wow" moment for me when using LLMs since the release of ChatGPT.

Somebody told me, to post videos of stuff I build on the weekend. So here is a simple demo of a much more powerful review agent I'm working on.

If your AI coding agents (cursor, claude code, devin) fail in a similar way to e.g. where new hires struggle, its likely your codebase fault and not the quality of the ever-improving LLM models. Looking at you custom time formats and 1-hour testruns.

Not just the best product, also the best support. Thanks! If you did not already switch to @linear, why?