Rijk van Zanten

Ozlo just launched the sequel to my most love/hate gadget of the last couple years: Sleepbuds. When they work, they're a game changer. Sleeping with white noise and having alarms just for me is amazing. However..

The original Sleepbuds were plagued with odd hardware issues. The amount of times I wanted to use them but found them out of battery because they either didn't charge or stayed on while in the case for some reason is infuriating.

I've admittedly spent an unhealthy amount on social media recently, and have to admit that X (fka Twitter) is wickedly more addictive than Bluesky. The constant barrage of (weird) hot takes and combativeness is pure nicotine. (Which don't get me wrong, is a bad thing)

Had a little /goal Codex experiment running over the weekend and am.. a bit underwhelmed. I gave it a relatively simple rename refactoring project, but it didn't get anywhere close to completion without blowing through my limits multiple times (on the Pro plan no less)

I do like the idea of it coordinating it's own subagents — as subagents do make the reliability and output quality a lot higher — but not at this insane cost. I kinda wish there was a way to run the "ultra" coordination of subagents against a medium or even low reasoning level

This year @directus.io has gotten 6× as many security reports as we did over the whole of 2025. This spike started in February, which is when Opus 4.6 launched (surprised? no) AI is great for improving security research, but the value has to be quality over quantity of reports.

Sleeper Mac app that I've been loving recently: Mole. Utility for cleanup, software uninstall, etc

Recently, I've been experimenting with a "rolling wave engineering" setup in codex. Instead of speccing out the whole project up front, I only spec out the broad strokes of the project, and then only go into detail one slice at a time.

So to resolve that, I've setup a new set of skills and agents to make the AI agents work a bit closer to how I do software development. The rough loop is: shape the project once up front, then loop over a prepare→implement→review→complete flow for each slice.

Recently, I've been experimenting with a "rolling wave engineering" setup in codex. Instead of speccing out the whole project up front, I only spec out the broad strokes of the project, and then only go into detail one slice at a time.

When starting the first "slice" of the project, I'd build that slice out in full detail, realize I forgot or have to reshape some of the project requirements, and then review and finish that one slice before moving on to the next.

Recently, I've been experimenting with a "rolling wave engineering" setup in codex. Instead of speccing out the whole project up front, I only spec out the broad strokes of the project, and then only go into detail one slice at a time.

The agent gets all bits of the project to abt 90% of the way there, but then I'm left with having to cleanup/review/tweak the last 10% of every step simultaneously across the whole project in a spec that might be completely wrong for what the project actually needs.

Recently, I've been experimenting with a "rolling wave engineering" setup in codex. Instead of speccing out the whole project up front, I only spec out the broad strokes of the project, and then only go into detail one slice at a time.

I've been trying Wispr Flow for a few weeks and learned something strange about myself: while it's undoubtedly faster than typing, I found myself going back to typing most of the time A) so it remains in my actual tone of voice, and B) because I hate murmuring to myself.

A is easy enough to resolve: I'm sure it can get smarter with learning from previous written text to match that more closely, but B) is a tough one. Reminds me of @reckless.bsky.social' matrix of wearable bullshit in which one axis is “usefulness” and the other is “fiddliness”.

Have you thought about changing to a sustaining engineering model focused on stability, security, reliability, and support yet? www.heroku.com/blog/an-upd...