Here’s my hot take: Build your own skills. Don’t waste your time bloating your context window with a bunch of junk you don’t need, made by someone who isn’t you and doesn’t have the same workflow as you.

Here’s my process:

• Get the agent to do a lot of stuff.
• Find a common process that has a lot of very particular context (rules, ceremony, etc.).
• Do it n times before you realize you should just make this process into a skill.
• Tell the agent to make the process into a skill.
• Use the skill the next time you do that thing.
• Polish the skill over time (add things that help and remove things that don’t help)
• Repeat as needed.

Yep, I wondered how I’d articulate this and you did it much better.

My approach with Claude Opus was basically: work with the agent and keep your eyes open. Now I have a really good harness where almost every single project quirk is addressed and the agent knows what to pay attention to, what it can glance and just move on, and what is a persistent danger that we must both always watch out for.

I reviewed those set of skills and generally I like them but they are not an end-all be-all. Same for a few others, including a guy I very much liked and worked with; they are good but nothing ground-breaking.

And the bespoke approach proved itself even more valuable after I started using Opus for my OSS work (deep into the pre-release work there, though it could still be months) – just find what the agent regularly trips on, add hooks and/or skills, monitor if it now does well, rinse and repeat.

I added these to my project:

It’s built for Claude, but I just copied over the markdown files in the agents and skills folder. They work great and have lots of eyes on it so anything crazy I imagine would get fixed by the swarm of people.

I use this with Opencode and Kimi K2.7 Code and GLM 5.2 works great!