// the find
foryourhealth111-pixel/Vibe-Skills
Intelligent Skill routing and workflow orchestration for AI agents — +21.12 pp reward, −29.6% tokens on SkillsBench with DeepSeekV4Flash-VE.
A skill-orchestration layer that sits in front of Claude Code, Codex, Cursor and similar tools: it scans whatever local 'Skills' (SKILL.md folders) you've installed, plans a multi-step task, and decides which skills to invoke and in what order. Aimed at people who've accumulated a pile of agent skills and want something to pick the right ones instead of relying on keyword-triggered matching.
The discovery design is legitimate: it builds a compact local index (name/description/boundaries) first and only loads full SKILL.md files for shortlisted candidates, which is a real token-saving pattern rather than dumping the whole skill library into context. It persists task state as plain JSON (module-work-plan.json, module-execution.json, delivery-acceptance-report.json), so a run can be resumed or audited after the fact instead of living only in a chat transcript. It also ships adapter configs for six different hosts (Claude Code, Codex, Cursor, Windsurf, OpenCode, generic) rather than hard-coding one tool's conventions.
The headline numbers (+21.12pp reward, -29.6% tokens) come from a benchmark repo maintained by the same author (foryourhealth111-pixel/vibeskills-benchmark) — there's no independent third party running this, so treat the precision of those stats skeptically. The package bundles 100+ unrelated domain skills (biopython, astropy, clinical trials, chembl, alpha-vantage, clinical-reports templates) inside what's supposed to be a thin routing layer; that's a strange amount of unrelated weight to pull in if you just want orchestration for your own skills. The actual orchestration logic (apps/vgo-cli) looks fairly small relative to the amount of README — localized SVG wordmarks, taglines, mermaid diagrams — which is more marketing surface than you'd expect for a dev tool. The account name and the 20-topic SEO tag list (vibe-coding, vibecoding, ai-scientist, context-engineering, automation, etc.) read like a repo tuned for GitHub trending/discovery rather than organic growth, so the star count is worth a grain of salt.