I tagged three releases over the last two days: agents v3.4.1, skills v2.5.0 and skills v2.5.1. Agents is my set of portable prompts and subagent roles for coding agents, and skills is a collection of agent skills for infrastructure, security and software engineering.
Most of this comes from one thing. Anthropic updated their skill-authoring best practices, so I checked my collection against them. It didn’t pass. Long reference files had no contents list, which matters because agents often only read the first 100 lines or so of a file. A few skills were over the recommended 500 lines. And most of the active skills predated the newer rules altogether.
So I rewrote them. The short version:
- Every SKILL.md is now under 500 lines, and lint enforces it. Long reference files get a generated contents list. The before-and-after chart shows the line counts.
- Workflows where order matters now have a checklist, with a clear step back when something fails. Tool menus became one default.
- The detail moved into reference files instead of being cut, so the skills got shorter without losing content.
Rewriting that many skills is an easy way to lose a safety rule without noticing. So every batch was reviewed by a different model, and it found real problems: dev-cycle had lost its release-authorization rules, and databases had lost a rule about bulk writes. Both are restored. I also ran a small A/B test on 8 of the skills, and the new versions scored about the same or a bit better (87.3 to 89.2 on average). It’s a small sample, so don’t read too much into it.
Separately, agents 3.4.1 tightens consolidate-agents-md. The backup and clearing steps used to be described loosely, so each run improvised the two most destructive parts. They’re now exact commands, and the backup is checked again right before anything gets deleted.
That’s the short version. The release pages on agents and skills have the full details. If a skill behaves worse for you after updating, please open an issue and tell me which one.