Vercel Agent Skills
vercel-labs/agent-skillsVercel's official 9-skill pack: React and Next.js performance rules, composition patterns, UI and writing review, React Native, deploy and cost audit.
27 internal agent skills inside the Next.js monorepo that help agents contribute to Next.js itself: PRs, backports, React syncs, tests and benchmarks.
with caveats
Well written and actively maintained, but these are internal contributor skills for the Next.js repo itself and do little for anyone building a Next.js app.
check first: makes network calls · installs packages → security details
This is the Next.js framework repository. It ships a set of agent skills under .agents/skills/, and they are written for agents working on the Next.js codebase, not for people building apps with Next.js. The sampled skills are marked metadata: internal: true.
The skills cover maintainer chores. create-pr walks the agent through staging only the relevant files, naming branches with an agent prefix such as claude/ or codex/, pushing, and opening a draft PR against canary with the repo's PR template. It also tells the agent not to write descriptions for fork PRs. backport-pr finds the merged commit for a PR number, branches from an LTS release branch, cherry-picks and opens the backport, and asks before touching git state if the target is unclear. Other skills cover syncing and vendoring React (react-sync, react-vendoring), feature flags, dead code elimination for edge builds (dce-edge), Rspack support, router act testing, test gating, PR status triage, stacked PRs (gh-stack) and error page work. authoring-skills explains how to write new skills and what belongs in AGENTS.md instead. A few skills include scripts: docs-diagrams renders SVG and PNG diagrams with headless Chrome, and sandbox-bench runs render and SSR benchmarks inside Vercel Sandbox VMs.
27 skills averaging about 10 KB of instructions, with reference files in a few skills and helper scripts in docs-diagrams and sandbox-bench; the repo's many test files cover the framework, not the skills.
Counted from the repository file tree. How we score depth.
This is the Next.js monorepo. Its internal agent skills drive git and gh to create branches, push, open draft PRs, merge stacks, reply to review threads and dispatch CI workflows when asked. The skills we could read only render diagrams locally and build React. Most of the repo, including the sandbox-bench scripts and the gh-stack extension, was not available for review, and the repo ships wasm and .node binaries.
gh extension install..agents/skills/gh-stack/SKILL.mdThe repository (34856 files) on Nov 1, 2026. Not a full review: 34160 of 34856 files were not downloaded (large repository; we reviewed the skill folders); it contains compiled binaries or archives we cannot read; we read 5 of 16 code files; it relies on code outside the repository: sandbox-bench skill scripts (.agents/skills/sandbox-bench/scripts/*.mjs, 11 files; full text not provided; they include '--sudo' and 'nohup' invocations); GitHub CLI extension github/gh-stack (installed by the gh-stack setup step); scripts/pr-status.js (used by pr-status-triage).
NVIDIA SkillSpector static scan: 40 pattern hits (Tool Misuse 21, MCP Rug Pull 11, Excessive Agency 4, Rogue Agent 3, Anti-Refusal 1). The rm hits are cleanup of output and temp files in render.sh and in the sandbox-bench scripts. We did not read the sandbox-bench scripts. The 'rm -rf /' hit appears to be a truncated match of 'rm -rf /tmp/warm-art'. The 'no warnings' anti-refusal hit describes expected test output, the npx hits are doc references to Next.js tooling, and the allowed-tools hit is an example in authoring documentation, not real frontmatter.
A review, not a certification. Skills change after we check them — read the files before installing.
MIT, commercial use allowed.
Not really. Most skills are single maintainer tasks like opening a PR or a backport that ask the user before risky git changes; only the benchmark scripts run long jobs, and they are not a general unattended loop.
Vercel's official 9-skill pack: React and Next.js performance rules, composition patterns, UI and writing review, React Native, deploy and cost audit.
25 engineering-lifecycle skills and 9 slash commands (spec, plan, build, test, review, ship) that make coding agents follow senior-engineer workflows.
A 15-skill development methodology that makes coding agents clarify the goal, write a plan, work through it with subagents and verify before calling it done.
The vercel/next.js repository contains 27 skills in .agents/skills/, but they are marked internal. They help agents work on the Next.js codebase itself: opening PRs, backporting fixes to LTS branches, syncing React, gating tests and running benchmarks.
Mostly no. They reference Next.js internals such as the canary branch, LTS release branches, config-shared.ts and the repo's PR template. If you want framework guidance for app work, look at Vercel's separate agent skills collection.
The scanner flagged many rm -rf and rm -f commands. Our review found they are temp file cleanup inside Vercel Sandbox VMs and generated-file cleanup in the diagram script. The one thing to watch is that create-pr tells the agent to retry blocked git and network commands with sandbox escalation.
It documents how the Next.js team writes skills. It lists the frontmatter fields they use (name, description, argument-hint, user-invocable, disable-model-invocation, allowed-tools, model, context, agent), how to write descriptions that auto-trigger reliably, and when content belongs in AGENTS.md instead of a skill.