agent-skills

“Been using AI coding agents daily for months now and recently discovered agensi.io, which is basically a marketplace for SKILL.md files. Bought a few, downloaded some free ones, and a handful have genuinely stuck in my rotation. Here are the 5 I keep coming back to: code-reviewer catches things I miss on my own PRs. Anti-patterns, style inconsistencies, security red flags. I run it before every push now and it's saved me from embarrassing commits more than once. env-doctor diagnoses broken dev e”

“env doctor alone sounds worth it lol. debugging setups is painful. same thing happens with automation flows once they grow.”

“built an agent skill that does something i haven't seen other skills do — instead of helping you write code, it helps you find problems in your specs*before*you write code. the idea: you write a concrete user scenario (persona, goal, failure modes), point it at your spec/design docs, and the skill walks through the scenario step by step, citing which spec governs each behavior and flagging gaps, conflicts, and ambiguities. been calling it "vibe testing" — like vibe coding but for the planning ph”

“As far as I can see this writeup is accurate I am still in the camp that it is best to still do SFT+RL for every new skill instead of just using a markdown file, however if you are going to just use a markdown file then the agents.md convention seems like a decent enough structure”

“Hey all you Claude power users. I've been working on leveling up my Claude use and that has been going pretty well, but I'm worried that I'm not approaching this the "right" way. I've read a lot about skills but still don't quite understand how to implement those, and I don't understand how they really differ from a custom agent. I do have a CLAUDE.md which i try to keep slim and focused. In the last few weeks I've gone from using the built in "Plan" agent and then moving to the "Agent" agent, a”

“What Happened Agent Skills dropped October 16th. I started building them immediately. Within two weeks, I had a cloudflare skill at 1,131 lines, a shadcn-ui skill at 850 lines, and a nextjs skill at 900 lines, chrome-devtools skill with >1,200 lines. My repo quickly got 400+ stars. But... Every time Claude Code activated multiple related skills, I'd see context window grows dramatically. Loading 5-7 skills meant 5,000-7,000 lines flooding the context window immediately. I thought this was just h”

“If anyone else found this AI slop obnoxious to try and read, here it is rewritten without all of the slop: My first approach to Agent Skills was creating massive 1,000-line files for tools like Cloudflare. The repo got popular fast, but activating multiple skills would flood the context window and grind everything to a halt. For two weeks, I was convinced dumping all the info upfront was the right move. The fix was to break each skill into three parts: a tiny metadata header, a short entry point”

“In general, it’s critical to never allow Claude to think it is writing “documentation” — which it will do by default for any markdown file. It is helpful to instruct it that what it is writing is not documentation, but instead a carefully architected file-based context cache for itself. It has a very strong and persistent bias toward bloating markdown files with trivial examples, obvious and countless examples of what not to do, and beginner/tutorial level fluff or generic and obvious guidance —”