For weeks the trending board has been a pile of agent skills, and that pile is still here. What changed this week is the layer forming on top of it. CL4R1T4S led with +524, then ponytail (+390), agent-skills (+370), apple/container (+339), and MiMo-Code (+324). The libraries that hand an agent more abilities are still trending, but the repos climbing fastest around them are now about restraint, taste, and policing what the skills do.
Start with the repo that frames the week. ponytail at #2 is a brand-new repo that took +390 stars this week on a single idea. Its pitch is one line: make your coding agent “think like the laziest senior dev in the room,” because the best code is the code you never wrote. Put it next to Agent-Reach at #9, which gives an agent eyes on Twitter, Reddit, YouTube, GitHub, Bilibili, and Xiaohongshu with no API fees. One repo wants the agent to see everything; the other wants it to do almost nothing. Both are in this week’s top ten. That gap is the actual story: the “more” tools and the “less” tools are now trending in the same breath.
The judgment cluster is wider than ponytail. taste-skill at #14 exists to stop an agent generating boring, generic output. shadcn/improve at #23 uses your most capable model to audit the codebase and write plans for cheaper models to run, which is a discipline patch dressed as a cost optimization. andrej-karpathy-skills at #25 is a single CLAUDE.md derived from one person’s notes on where LLM coding goes wrong. None of these add a capability. They each try to make the agent exercise some judgment it does not have by default.
Then there is the part that says the category got big enough to attack. SkillSpector at #18 is NVIDIA shipping a security scanner for agent skills, built to flag vulnerabilities and malicious patterns inside the skill files themselves. A vendor that size writing a scanner for an artifact class is a signal about that class: skills are now common enough, and trusted enough by default, to be a real attack surface. cc-switch at #13 sits next to it for a quieter reason, managing Claude Code, Codex, OpenCode, OpenClaw, Gemini CLI, and Hermes Agent from one desktop app because developers now run too many agent CLIs to track by hand.
The infrastructure repos held their ground without any of this framing. apple/container at #4 runs Linux containers through lightweight virtual machines on Apple silicon, and it keeps showing up because every one of these agents needs a local box to run in. iptv-org/iptv at #11 is a public M3U playlist index with no model and no agent, trending on plain usefulness. microsoft/markitdown at #17 converts documents to Markdown, the same pre-processing step that keeps feeding the whole stack. These three would have trended in 2022 for the same reasons they trend now.
The honest boundary has not moved. These are rolling seven-day star deltas, and “skill” still carries enough naming momentum in June 2026 that some entries are one-file experiments riding the fashion rather than tools anyone will run in three months. The week’s #1, CL4R1T4S, is leaked system-prompt material, so its rank is attention and curiosity, not adoption. But that curiosity is consistent with everything else here. People want to read how the popular agents are configured, then ship their own configuration as skills, then audit the skills they install. New models barely registered this week. What climbed instead were the parts that decide how a model should behave, plus the first tools built to keep that layer honest.