The State of Agent Skills: What I Found When I Asked 'Should We Build a Skill That Writes Skills?'

It started with a simple question. I was looking at obra's writing-skills, a popular open-source guide for authoring AI agent skills, and I wondered: do we need something like this for our own skill library?

The answer turned out to be less interesting than what I found along the way.

The ecosystem grew up while nobody was watching

Three months ago, "Agent Skills" was a newly announced open format from Anthropic. Today, it's an ecosystem. There are tens of thousands of skills on registries like skills.sh; the format works across dozens of platforms (Claude Code, Cursor, GitHub Copilot, OpenAI Codex, Gemini CLI, and others).

Then there's the other standard: AGENTS.md, the "README for AI agents," with formal stewardship and a v1.1 proposal that positions it as complementary to SKILL.md: always-on project guidance vs modular, on-demand capabilities.

Two standards, one ecosystem. They are converging more than colliding. That's the good news.

The quality problem

Piyush Vyas audited 53 published agent skills from Anthropic, OpenAI, and Vercel. Four failed to parse entirely; three of those four were Anthropic's own published skills, broken by missing file references. Of the rest, most lacked error handling; agents get no recovery path when instructions fail.

The organization that created the Agent Skills specification has shipped skills that don't fully comply with it. That's not a dunk; it's a signal: skills are snapshots; they don't update themselves.

Staleness at scale

Then came Chris Williams at Vercel, who wrote with unusual honesty about maintaining dozens of skills across many products: release cadences differ, packages get renamed, and agents confidently follow stale install lines. He's proposed tooling direction (skill-versions, quality checks) so "agent knowledge" can be reasoned about more like dependencies.

We've solved drift for code. For agent knowledge, we're still early, but skillscheck.ai and related tools are pointing at the same problem.

The gap that matters

Should we build a skill that writes skills? Plenty of generic guides exist: obra's methodology, Minko Gechev's skills-best-practices, Anthropic's guide, the agentskills.io spec. None of them encode our routing tables, handoffs, aliases, and approval gates.

The gap isn't "how to write a SKILL.md." It's "how to write one that works in your system."

What we're doing

We created 829-create-skill, a Lab829 skill that teaches agents our conventions on top of the spec, and repo-level AGENTS.md for ~/.agents so cold-starting agents get orientation without spelunking. We also have a skill-improve with sub-workflows like audit and fix that we sporadically run to benchmark. But most importantly, ALL of the skills that we create are now getting monitored, and ALL of them enforce cross-references that check if, for example, a commit is actually working as expected; if it is not, the skill will flag the issues so we can audit again and fix the issue.

Feels like a lot of work but continuous and proactive guard from the skills is helping us deliver and ship code faster, meeting proper standards and having fewer surprises when code goes to production.

If you're standardizing how agents use skills and project context, we're happy to help. Lab829 works with teams on agent tooling, product delivery, and the glue between spec and practice.