Matthias Lohscheidt

It usually happens when I forget to clear or don’t take the time to properly manage the context. "Just
one more thing …"

Is it just me or does reaching auto compaction with Claude Code feel like fail?

It probably would work, but requires the agent to compare the content of this file with whatever change it did, every time. Could solve the drift issue, but still extra work for questionable outcome. But sure, fresh docs can have value on their own. Maybe a human comes along and wants to read them

Delete your CLAUDE.md (and your AGENT.md too) YouTube video by Theo - t3․gg

The data seems to suggest that AgentMD files are not helping most of the time. Agents are really good in getting the context they need from the code and the old dilemma of outdated docs is true with AgentMDs as well.

www.youtube.com/watch?v=GcNu...

Screenshot of CLAUDE.md file: # CLAUDE.md The role of this file is to describe common mistakes and confusion points that agents might encounter as they work in this project. If you ever encounter something in the project that surprises you, please alert the developer working with you and indicate that this is the case in the Claude.md file to help prevent future agents from having the same issue.

When you just opened a PR with a finely composed ClaudeMD and then @t3.gg comes along with his proposal:

@eslint.org is even more valuable in times of agentic coding. A very fast automated feedback loop.

Best part: if weird patterns emerge that are specific to your project, let AI write custom linting rules to prevent them in the future.

And be careful that it doesn't disable them too lightly…

Biggest moat for SaaS against vibe-coded solutions: all the laws out there regulating digital products and services. At least when public facing.

(CRA, DSA, GDPR, to name just a few EU-based ones)

This line in my personal preferences in Claude is gold. I got honest pushback on some assumptions in my conversations with Claude recently:

> When I present flawed ideas, I like to be challenged and offered alternative approaches.

So true, thanks for sharing ❤️
I have two daughters (5, 2) and they are the most important human beings in the world to me. And as you said it's a rollercoaster between feeling grateful and exhausted.
As soon as they start to play together, it gets better. Plus more fights 😅

CEOs of big tech companies who the Trump administration listens to have decided to ignore the slow descent into fascism. www.resistandunsubscribe.com

Using Wispr Flow since two weeks now. Funny to see how my writing brain seems not to be the same as my speaking brain, and how I need to "merge" them to effectively use a speech-to-text tool like this.

Form follows function. I like it :)

I use zip lock bags to keep them dry*ish.

I think I need an AI to help me understand my Cursor invoices.

Field Combine labels, controls, and help text to compose accessible form fields and grouped inputs.

The new @shadcn.com/ui Field component is a delight. Really well thought through: ui.shadcn.com/docs/compone...

I feel you! Curious how effective this is. I use Spark's block feature, which allows me to block any further emails from a specific sender (because they usually send a ton of follow-ups before they give up).

Wow, that‘s huge, @amann.work! This could also be a path towards optimized message bundling per component tree, right?

Code sample that demonstrates useExtracted

🌐 next-intl 4.5: useExtracted

→ For humans & AI
→ Next.js-first by design
→ Full Turbopack support
→ SWC-based parsing

Fix those flaky tests. Now! They are a pain in the neck, but very easy to ignore. Here's why you shouldn't.

Still though: www.mattiloh.de/posts/fix-fl...

Love this small-but-mighty update from @circleci.com: circleci.com/changelog/au...

The number of times I’ve fired up a Node REPL just to run crypto.randomUUID()…

Meanwhile, uuidgen has been on my mac this whole time. 🤦‍♂️

It’s a slick combo: natural language in, deterministic actions out. And it's framework-agnostic. I integrated it into and old React + Redux setup. I guess with React Server Actions things could be even simpler.

AI Command Bar - Turn your user's words into actions | Product Hunt Imagine Spotlight with superpowers... That is what Stepsailor's AI Command Bar is aiming for. Let your user describe what they want to do, and the command bar will automatically run the required actio...

Got early access to this and plugged it into an internal admin tool. Super easy integration and instantly killed a ton of repetitive grunt work. Great job @pascal-lohscheidt.com!

www.producthunt.com/posts/ai-com...

Yes! It’s 20y since I watched it, but I remember how I was fascinated by its aesthetics. If I remember correctly, it was shot on 70mm film to be able to capture the many little details in the numerous wide shots.

Lego Brio train by mattiloh | Download free STL model | Printables.com A little toy train, connecting two fantastic universes: the creativity of Lego bricks and the classic fun of Brio tracks | Download free 3D printable STL models

Enjoy! It's so much fun and it's really a great tool to fix / improve little things in the house. And there's also things like this: www.printables.com/model/109151...

I'll be in Amsterdam later this month for DevWorld Conf, where I'll be giving a talk all about Node.js in the past, present & future.

🎫 I have 2 free tickets to give- for a chance to win, like + share this post. I'll announce winners on Friday!

https://buff.ly/42MeWGV

#devworld

User deletion from Mailchimp lists via API is a mess. A regular delete only archives the user (not #GDPR compliant). A permanent delete means you can’t re-add them later if they sign up again – unless you do it manually through Mailchimp’s UI.

Another day, another dry kiss from Yagni. #CleanCode

A screenshot of a conversation with GitHub Copilot. The user asks if multiple events, like pull_request and workflow_run, in a GitHub Actions workflow are connected using boolean logic, specifically whether all triggers must occur or if just one suffices. Copilot responds by explaining that the events are connected with an OR logic, meaning the workflow will trigger if any one of the events happens. The interface includes a prompt bar labeled "Ask Copilot" with icons for attachments and actions.

I rarely used GitHub Actions, mostly CircleCI in the past. Just reviewing some GH actions and the integrated Copilot AI is proving to be helpful for gaining more insights without the need to context switch.

Cool! We wrote a very similar plugin last week. We didn’t open source it yet, since it was just a quick internal implementation for a prototype. I guess it’s tricky to make a generic linting rule that works for everyone. E.g. we needed to check for unwrapped t() function invocations (next-intl).

Reposting. Still looking for help on this.