On a single dev machine running Claude Code (Anthropic's AI coding assistant) and Codex (OpenAI's equivalent) side by side, 13 copies of what should be identical AI playbooks have silently drifted apart — the largest gap between any two versions is 58 lines, with zero error warnings. We care about this because enterprise AI is moving from "install one tool" to "run several at once," and the same kind of version mess scales multiplicatively.
What this is
Here, a "skill" is best understood as a standard operating procedure (SOP) fed to the AI: natural-language instructions that tell it, "when you hit this situation, follow these steps." AI coding assistants like Claude Code and Codex — collectively called Agents (AI programs that can autonomously understand and execute multi-step tasks) — both support loading skills.
The trouble starts when you run them together. On the developer's machine, 50+ skills come from five different sources (the official plugin marketplace, the company's internal marketplace, global npx installs, manual git clones, and hand-rolled scripts), each with its own versioning scheme. In the ~/.claude/skills and ~/.agents/skills directories, 13 skills share the same name but have already diverged in content — whichever copy the model reads, it executes, and the outcome is unpredictable. Two more hand-written skills are still calling APIs deprecated six months back: format valid, triggers firing — just outdated.
Beyond cost, there's trigger overhead: 21 skills alone keep roughly 3,219 tokens (the smallest unit of text an AI processes — about one Chinese character) permanently loaded in their trigger descriptions, whether you need them or not. Four image-generation skills do nearly identical jobs with heavily overlapping trigger words, so which one the model picks is a coin flip.
Industry view
The open-source community moved fast. On GitHub, vercel-labs/skills (33,000 stars) and xingkongliang/skills-manager (5,600 stars) share the same idea: keep one master copy, drop symlinks (shortcuts pointing to the same file) in every agent directory. The first behaves like npm (Node.js's package manager); the second acts like a local skill library plus a desktop client.
But the sober side: neither tool solves trigger conflicts or skill effectiveness evaluation. In other words, "installed" is solved — "which one will fire and how well it works afterward" is still blank. The original author's stopgap is Claude Code's built-in /skill-doctor and claude plugin eval. In short, no industry-grade answer exists yet.
A deeper read: once enterprise IT starts stitching multiple AI tools together, this kind of "master copy drift" will spread from the developer's machine to the whole organization. Today it's the developer's pain point; next year it may be the CIO's.
Impact on regular people
For enterprise IT: if the company plans to roll out two or more AI tools at once, planning "who owns the master copy, how it's synced, who audits it" up front will save a lot more than firefighting later.
For individual careers: people who can author and maintain AI SOPs (skills) well will be rarer than pure prompt-writers as enterprise AI lands — the prompt bar is being rapidly flattened.
For the consumer market: not much for the moment. Current solutions target developers and enterprise IT; ordinary consumers are still a year or two out.