Name a Skill verify and Claude Code Will Run It Before Committing

Claude Code shipped six releases in the eight days from September 24 to October 1, 2026. Most of it is the usual plumbing — plugin installs, MCP reconnects, screen reader fixes. But buried in v2.1.286, released September 30, is a one-line change that quietly turns a skill's folder name into an interface:

No frontmatter field opts you in. No settings key. The name of the directory is the trigger. That is a meaningful shift for anyone who writes or installs skills, and it arrived alongside two more name-sensitive behaviors and a new command that audits your skills for rot.

What changed on September 30

Before this release, Claude Code's commit behavior was driven by whatever you wrote in CLAUDE.md. If you wanted tests run before a commit, you said so in prose and hoped the instruction survived the context window. Now there is a canonical place to put that work: a skill named verify. Claude is told to run it as the last step before a commit, with a sensible carve-out for docs-only and tests-only commits where a full build adds nothing.

Read the wording carefully, because the distinction matters: Claude is told to run it. This is model guidance, not a deterministic gate. A PreToolUse hook that shells out to your test command will block a commit; a verify skill makes it overwhelmingly likely Claude runs the checks and reports honestly. For most teams that is the better trade — the skill can reason about what changed and pick the right subset of checks, which a hook cannot. For compliance-grade enforcement, keep the hook and the CI gate too.

Where the verify skill lives

Claude Code looks for it in the normal skill locations: project skills at .claude/skills/verify/SKILL.md and personal skills at ~/.claude/skills/verify/SKILL.md. Project wins for team-wide gates — commit it and every agent in the repo follows the same recipe.

That path may already exist in your repo without you writing it. Claude Code's bundled /verify skill — the one that builds and runs your app to confirm a change does what it should, without falling back to tests or type checks — records its own recipe. When it has to figure out how to build and drive your app from scratch, it writes what worked to .claude/skills/verify/SKILL.md at the repo root (or in the touched package directory in a monorepo), and at the repo root that recorded skill replaces the bundled one. So the September 30 change plugs straight into a file the tool has been generating since v2.1.200.

---
name: verify
description: >
  Verify this repo before a commit. Runs unit tests, the typecheck,
  and the production build, in that order, and reports the first
  failure with its raw output. Use before committing code changes.
---

Run these in order and stop at the first failure:

1. `pnpm test` — unit tests across all packages
2. `pnpm typecheck` — no new type errors
3. `pnpm build` — the build must succeed

Report the exact failing command and the last 20 lines of its
output. Never summarize a failure as "mostly passing". If every
step passes, say so in one line and nothing more.

If you would rather install a vetted version than write one, the registry has close fits. verification-before-completion is built for exactly this moment — it fires when an agent is about to claim work is complete, fixed, or passing, and requires running verification commands and confirming output before any success claim. autoreview runs a pre-commit code review with Codex by default. And the 561-install code-review skill is the obvious companion for the review half of the gate.

The names that silently don't load

The flip side of names carrying meaning is names being taken. Two skill folder names will not load at all, per the Agent Skills documentation:

• synced — in any capitalization. Claude Code uses ~/.claude/skills/synced/ for skills downloaded from claude.ai and skips a skill you author at that name in the enterprise, personal, and project locations.
• anthropic-skills — outside a plugin, a skill folder or command file named anthropic-skills or starting with anthropic-skills: does not load.

That list was briefly longer. v2.1.282 on September 24 reserved both anthropic-skills and claude-ai as namespaces where skill folders, command files, and workflow commands no longer load — and also stopped MCP servers configured under either name from listing skills or prompts at all. One day later, v2.1.283 reverted the claude-ai half: skills, commands, workflows, and MCP server prompts under that name load again, and Skill(claude-ai:*) rules went back to being ordinary prefix rules.

/doctor prompt-audit reads your skills

The same v2.1.283 release added /doctor prompt-audit (also available as /checkup prompt-audit), which audits your CLAUDE.md files, skills, agents, and commands for prompting patterns written for older models. A refinement in the same release pushes stale paths, stale commands, and contradicting instruction files to the top of the report, while keeping the thinking keywords Claude Code actually documents.

This is the first first-party tool that treats skill decay as a real maintenance category. It matters most for installed skills, not authored ones. A skill published against a weaker model tends to be padded with defensive scaffolding — repeated warnings, step-by-step hand-holding, instructions to re-read the file. On a current model that padding costs tokens and can actively steer the run worse than saying nothing. Running the audit across your whole ~/.claude/skills/ directory is a cheap way to find out which of your installs are carrying dead weight.

There is a longer-standing variant worth knowing: the bundled /claude-api skill has had its own prompt-audit subcommand since v2.1.221, which flags instructions written for older models in your prompts, skills, and tool descriptions and proposes fixes as a diff. If you ship an LLM-backed product, that one audits your runtime prompts rather than your Claude Code configuration.

For the curation half of the problem, the registry has review-a-skill, which exists to decide whether a skill belongs in a library at all — preview the content, check the frontmatter, validate the structure, then keep, curate, or remove. Pair it with constraint-driven-development, which writes your quality bar down as a contract so agents can't quietly lower it, and test-gap-audit for a read-only sweep of missing, weak, or stale coverage.

Skill() rules now match aliases

The third name-sensitive change in v2.1.283 is for anyone governing what an org's agents may invoke. Skill(anthropic-skills:<name>) deny rules now also block that skill when Claude Desktop delivers it as a plugin, and Skill(skill:<name>) denies now match the skill's alias and display name, not just its canonical one. If your deny list was written against a single spelling, it was dodgeable by whichever other spelling the host happened to use. It isn't now.

A smaller fix in the same release will save someone an hour: when a skill belongs to a plugin that failed to load, the Skill tool now says the plugin could not be loaded instead of reporting the skill as uninstalled. And v2.1.284 on September 28 fixed repositories added mid-session in cloud and SDK sessions not loading their skills and plugins at all. If you run agents in cloud sessions against multi-repo checkouts, that one is worth the upgrade on its own.

If permission boundaries for coding agents are new to you, ai-coding-agent-guardrails covers permission boundaries, secret protection, and review gates across Claude Code, Cursor, Codex, and Copilot, and common-agent-guardrails defines deterministic guardrails for tool calls — protected paths, test-file locations, and the rest.

What to do this week

Four things, in order of how much they'll save you. First, check whether you already have a verify skill and whether what it does is what you'd want run before a commit. Second, if you don't have one, write it or install one — this is the cheapest pre-commit gate Claude Code has ever offered. Third, rename anything called synced or anthropic-skills; it has been silently ignored and you may not have noticed. Fourth, run the audit.

# 1. What skills do you actually have, and is one of them named verify?
ls ~/.claude/skills/ .claude/skills/

# 2. Install a verification skill if you'd rather not author one
npx @skills-hub-ai/cli install openrig-verification-before-completion

# 3. Audit CLAUDE.md, skills, agents and commands for stale prompting
#    (run inside Claude Code, needs v2.1.283 or later)
/doctor prompt-audit

The larger pattern is worth naming. Agent Skills started as a pure discovery format: the description field told the model when to reach for a skill, and the folder name was just an address. Claude Code has now loaded that address with behavior in three directions at once — one name triggers a lifecycle step, two names are reserved, and all names are matched by permission rules including their aliases. Skill names are becoming shared vocabulary the way Makefile and test are, and the ecosystem will keep adding entries. Pick names for the job, not the vendor, and you'll keep landing on the right side of it.

# Browse verification and review skills on skills-hub.ai
npx @skills-hub-ai/cli browse --category review
npx @skills-hub-ai/cli browse --category qa

Full details are in the Claude Code changelog and the bundled skills reference. For more on the authoring side, see building your first skill or browse the review category and the Anthropic source on skills-hub.ai.