Let One Skill Route All Your Local Skills With VibeSkills

Install enough skills and a new problem appears: you forget what you have, and your agent never checks. VibeSkills, a 4k-star general-purpose skill, sits above that library — it discovers the skills already installed on your machine, assigns the relevant ones to each part of a task, and verifies the combined result before delivery.

Why This Skill Matters

Passive skill triggering reacts to a few obvious words; VibeSkills splits the whole task first. Its README describes a five-stage process: confirm the requirement (the process stops until you approve the goal, constraints, material, and expected delivery), recommend an effort level — L for manageable multi-step work handled in order, XL for larger work where up to two non-conflicting parts run at the same time — organize skills by reviewing your local skill folders, execute and record (code tasks can run a TDD loop), and finally check the actual result against every planned item. Incomplete, failed, or blocked required work prevents final acceptance.

The README reports benchmark numbers for this approach: on its adapted large-scale SkillsBench setting — every task evaluated with all 195 specialized skills installed — vibeskills v4.1.0 raised mean verifier reward by 21.12 percentage points over the baseline (DeepSeekV4Flash-VE with OpenHands), while cutting total tokens by 29.6% and tool calls by 33.1%. The README's own analysis attributes the gain to selecting only a few truly relevant skills per task, not invoking more of them.

The token question has a documented answer too: discovery and index generation happen locally against compact extracted information — name, description, intended use, boundaries — and only retained candidates are read as complete SKILL.md files. The README notes this overhead is not zero; more candidates and longer documents cost more context.

Installation

The public install path starts from a published release zip, not a repository checkout: download vibe-skills-4.1.0-public.zip from the releases page, verify its SHA-256 digest against the one shown on the release page, and extract it outside the managed skills directory. Then run the wrappers from the extracted folder:

bash ./install.sh --skills-dir "$HOME/.agents/skills"
bash ./check.sh --skills-dir "$HOME/.agents/skills"

The default directory is ~/.agents/skills; pass a different one — ~/.claude/skills, ~/.codex/skills — and the same runtime installs there. The installer writes only to <SkillsDir>/vibe: it does not edit Codex, Claude, or Agents settings, does not write system prompts or command wrappers, and does not automatically install or recommend the chrome, chrome-devtools, playwright, context7, or claude-flow MCPs. check proves the install is present, not that a task ran. Removal is deleting <SkillsDir>/vibe.

Real Workflow: A Multi-Deliverable Task Through $vibe

Invoke VibeSkills through your application's skills entry with $vibe, /vibe, or the syntax it provides. Then hand it the whole task, deliverables included:

$vibe 用公开数据完成一个可复现的分类实验：交付数据审计、统计评审、4 张结果图、一份科学报告，以及一个 7 页的组会幻灯片。

The README's own case follows exactly this shape: with more than 100 skills configured on the host, VibeSkills reviewed the candidates and their SKILL.md files, selected 7, arranged the work into 5 groups and 10 work units — environment, data audit, modeling, statistical review, figures, report, slides — and ran 17 checks across the deliverables. The task passed acceptance at 10/10 units completed and 17/17 checks passed. Your task record keeps the approved requirement, plan, execution progress, and final check in one place, so a later session can continue from saved progress.

Tips

• Add skill folders beyond the shared directory through ~/.vibeskills/skill-roots.json (user level) or <workspace>/.vibeskills/skill-roots.json (workspace level); per the README, a skill needs a readable SKILL.md, a name that does not conflict with another Skill, and a clear fit for the current work before it can be selected.
• Finding a skill means it can be considered, not that it took part — agent_skill_organization stores the planned assignment, module_assignments the actual one.
• Confirm the requirement honestly: the process deliberately stops at stage one until you approve, and the plan and final check are only as good as that approved basis.
• A successful installation is not a completed task, and a task record is not a passed check — the README states both distinctions explicitly.

When Not to Use This

If you run fewer skills than you can hold in your head and your agent already picks the right one, an orchestration layer adds a planning stage you do not need. And it is not an installer of capabilities: VibeSkills selects from what is already on disk, so capabilities that are not installed are simply unavailable to the plan.

See the leaderboard for more skills.