skill.md Becomes the Cross-Vendor Agent Extension Standard
Claude Code's third-party plugin count crossed 100 this week, with six skills emerging as the production baseline for commercial AI work. On the same day, Google's edge AI team revealed that Gemma 4 independently adopted the identical skill.md progressive-disclosure pattern for on-device agent capabilities. GitHub Trending confirms the escape velocity: a cross-tool Skills Manager now synchronizes agent skills across 15+ AI coding tools simultaneously.
What the Source Actually Says
Nate Herk, who audited Claude Code's plugin ecosystem across 400+ hours of client work, distilled 100+ available skills to six production picks: Skill Creator (Anthropic's official meta-skill — generates skills from plain-English descriptions), Superpowers (150K+ GitHub stars, enforces plan-before-code workflow), GSD (sub-agent context isolation against context rot), /ultra review (parallel cloud-sandboxed reviewers that surface only independently-verified bugs, co-launched with Opus 4.7 at $5–20/run after three free runs), Context Mode (sandbox routing that shrinks 56 KB Playwright snapshots to 299 bytes, extending sessions from 30 minutes to 3 hours), and ClaudeMem (cross-session SQLite + vector memory with 3-layer retrieval, claimed 10× token savings vs startup dumps).
The cross-vendor corroboration comes from Google's Cormac Brick (tech lead, Google AI Edge), who confirmed that Gemma 4 E2B/E4B models use an architecturally identical structure: each skill is a skill.md file with a one-line description visible by default and full instructions loaded on demand. Brick credited "a pattern you may have seen elsewhere" — a direct reference to Anthropic's approach — and noted that Google built 80+ AI Edge Gallery skills this way, "more than half vibe-coded" using Gemini CLI and Claude Code. On GitHub Trending the same day, xingkongliang's Skills Manager (Tauri 2, Rust) surfaced as a unified skill OS for Cursor, Codex, Gemini CLI, GitHub Copilot, Windsurf, and 10 more tools. A separate entry, the iOS Simulator Skill for Claude Code, reported a 100% eval pass rate versus 46% without the skill and a 96% token reduction — quantifying what mature skill design delivers in production.
Strategic Take
Two of the largest AI vendors independently converged on the same skill.md progressive-disclosure architecture — the canonical signal for a de facto standard forming. Teams building agentic systems should adopt this pattern for their own extensions now: one-line description in context by default, full spec loaded on demand. The cross-tool infrastructure is already live; skills written today travel across vendor boundaries tomorrow.
