A while back, I wrote you-dont-know-yet, a skill for coding agents. It was inspired by the i-have-adhd skill, which I found pretty useful. While I liked how it made responses more direct, I wasn’t a fan of its five-item cap on lists. At the same time, I was getting increasingly annoyed at models (especially Opus 5 a while back, those who know will know) being overly confident, like telling me they’d fixed something when all they’d done was change a few lines of code, without actually checking whether the fix worked. So I wrote my own version, keeping what I liked and adding a clearer distinction between what had been changed, what had been verified, and what was still unknown.
Fast forward a few months, and with the arrival of GPT-6, the Claude 5.5 series (Opus, Sonnet, and more recently Haiku), and a bunch of other increasingly capable models, I’ve deleted a good number of those skills. I found myself needing fewer instructions to get the behaviour I wanted, with some skills even becoming more of a burden than a help as the model limitations they were written to work around became less of an issue. Rather than continuing to write skill after skill and stuffing the model’s context with more instructions, I’ve started using the coding harness in a more raw way, letting the model figure out how to approach things without so much hand-holding.
That’s not to say I’ve stopped using skills altogether. I still find hooks incredibly useful, particularly for enforcing certain checks and actions without relying on the model to remember them. Certain skills, like enforcing a consistent PR format, are also worth keeping. They’re useful for specific, repeatable tasks, rather than trying to dictate or steer how the model thinks.
This whole experience got me thinking about how much effort we spend keeping up with AI. In the last four years, I’ve gone from using OpenAI’s free API credits (I remember getting $10 as a new user, with an expiry date) and the old Ada, Babbage, and Davinci models to write joke bots on Telegram with a friend for a hackathon we turned up to mainly for the swag, to experimenting with basic RAG, and now to the age of agents. I’m genuinely excited about what’s to come. With another model, framework, tool, or supposedly better way of doing things coming out every other month, week, or even day, though, it’s easy to feel like you’re falling behind if you’re not constantly reading up on everything.
I’ve tried subscribing to newsletters and following “tech influencers”, but I don’t particularly enjoy keeping up that way. I’d much rather come across something interesting on X or Reddit, follow a rabbit hole, and try it out for myself. Sometimes it turns out to be genuinely useful. Most of the time, though, I end up spending hours exploring stuff I’ll probably never touch again (just look at my GitHub repo count versus the number of things I’ve actually produced that are useful, lol).
I guess I’m becoming less concerned about whether I’m keeping up with everything new, and more interested in just trying random things I find interesting, keeping what works, and moving on when something better comes along. I’d much rather let exploration be driven by genuine interest and curiosity than turn it into another item on my to-do list, or yet another video saved to “Watch Later” that I’ll probably never return to.