Your Claude workforce — 44 specialist skills across 7 departments, behind one router. Ship code, design interfaces, rank in search, model the numbers, run ops, read the contract.
Most skill collections do one of two things badly: they cover a single domain and leave you switching tools for everything else, or they dump a hundred skills into your session and quietly wreck the context budget every other skill depends on. OmniSkill does neither. It gives you a real specialist for each of the seven functions a small company actually runs on, routed from a single entry point — and it costs your skill listing 8 entries, not 44. Invoke it yourself with /omniskill, or let it offer itself when you start a new project. It never activates without your say-so.
Built for Claude Code (Windows / macOS / Linux).
| # | Department | Covers | Skills |
|---|---|---|---|
| 01 | Developers | Ship code faster, from scaffold to QA | grill-me · superpowers · context7 · mcp-builder · skill-creator · webapp-testing · claude-mem |
| 02 | Design | UI that never looks templated | ui-ux-pro-max · frontend-design · web-artifacts · motion-ui · canvas-design · algorithmic-art · slack-gif |
| 03 | Marketing | Copy, SEO and ads that convert | seo-audit · programmatic-seo · ai-seo · cro · ad-creative · mktg-psychology |
| 04 | Social & Content | Feed the algorithm on autopilot | social · copywriting · content-strategy · video · pillar-content · email-sequences |
| 05 | Finance | Model the numbers before you spend | dcf-model · 3-statements · lbo-model · comps-analysis · pricing · pitch-deck |
| 06 | Operations | Run the business like a machine | sop-builder · incident-postmortem · business-case · launch-runbook · internal-comms · xlsx |
| 07 | Legal | Read the fine print for you | contract-review · nda-triage · legal-risk · compliance · docx · sql-queries |
Full map with one-liners: references/registry.md.
- Seven departments, one door —
/omniskillreads the job, recommends a crew, and loads only that specialist. Real projects cross crews, so it sequences them (Finance → Design → Marketing → Legal → Ops) and loads each only when it gets there. - Costs 8 skill-listing entries, not 44 — the routers carry descriptions; all 44 leaves are
disable-model-invocation: true. They cost zero listing budget and stay fully invocable by name. On a machine already running several plugins, this is the difference between adding a workforce and degrading every skill you own. - Playbooks, not prompt-shapers — 31 skills are full working methods. DCF enforces unlevered means unlevered and sanity-bands the terminal value. LBO attributes IRR across deleveraging, EBITDA growth, and multiple expansion. Contract review ranks findings by severity against deal size. These are methods you'd actually follow.
- Nothing mature gets re-implemented — 31 skills are written here; the other 13 wrap real upstream work. Ten of those ride along in the box, vendored unmodified with their licenses intact, so a fresh machine gets the whole workforce from one install and fixes still go upstream.
- Offers itself, never assumes — a hook notices new-project intent and suggests. Claude then asks. Every activation goes through a question you answer.
- Safety is not an upsell — Finance and Legal state plainly that they are analysis, not licensed advice. Marketing refuses to build dark patterns regardless of measured lift.
claude plugin marketplace add IPedrax/OmniSkillclaude plugin install omniskill@omniskill
claude plugin marketplace add ./OmniSkill && claude plugin install omniskill@omniskillThat single install is the whole workforce. Ten upstream skills ship inside it — superpowers (all 14), grill-me, motion-ui, ui-ux-pro-max, canvas-design (with its 54 OFL fonts), web-artifacts, slack-gif, mcp-builder, frontend-design, and skill-creator. Nothing to chase down.
| Skill | Why | What to do |
|---|---|---|
xlsx, docx |
Source-available and proprietary — cannot be redistributed in an MIT plugin | Ship with Claude Code in most setups. Otherwise claude plugin marketplace add anthropics/skills |
context7 |
An MCP server, not skill content | The adapter offers to set it up when you first need it |
Restart Claude Code after installing. Hooks load at session start, so new-project detection stays dormant until you do.
Requires Claude Code ≥ 2.1.71, the version this is built and tested against.
Prefer to let Claude install it? Paste this into a new chat: "Install this Claude Code plugin for me from https://github.com/IPedrax/OmniSkill and walk me through anything you need."
Just ask, and let it route:
- "build me a landing page for a new SaaS"
- "what is this company worth"
- "review this contract before I sign"
- "why isn't this page ranking"
- "write the go-live runbook, we cut over Friday"
Or go straight to what you want:
/omniskill pick a crew for what you are doing
/omniskill:finance jump to a department
/omniskill:dcf-model jump to a specialist
All 44 leaves are directly invocable by name.
Detect intent → offer the crew → you approve → load the department → load the one specialist → work.
When you start a new project, a UserPromptSubmit hook notices and injects a suggestion — nothing more. It cannot load a skill and it cannot act; Claude asks you via a real question and waits. It fires at most once per session, only on wording that implies building something new ("build me a new app", "from scratch"), and never on existing-codebase work ("fix this bug", "refactor this"). Vague requests only trigger it in a directory that is genuinely empty.
To turn it off entirely, set OMNISKILL_AUTO_SUGGEST=false in your environment, or in the env block of ~/.claude/settings.json:
{ "env": { "OMNISKILL_AUTO_SUGGEST": "false" } }
Claude Code loads every skill's description into context, capped at roughly 1% of the context window. On a machine with several plugins installed that budget is already tight, and when it overflows Claude Code starts dropping descriptions — from every skill you have, not just the new ones.
So OmniSkill spends 8 entries and no more. The routers carry descriptions; all 44 leaves set disable-model-invocation: true, which costs zero listing budget while leaving them fully invocable by name. Routing reaches them by reading their SKILL.md directly.
The trade: Claude won't spontaneously reach for dcf-model — it gets there through /omniskill or a department router. Want a specific leaf auto-triggering? Delete one line from its frontmatter.
node scripts/check-registry.mjsChecks the tree against the registry, confirms all 44 leaves exist, and asserts exactly 8 skills are model-invocable — the context-budget guarantee. It fails loudly if the architecture ever regresses.
node hooks/detect-new-project.mjs --selftestChecks that new-project wording fires, existing-codebase wording stays silent, the output shape is right, and the once-per-session dedupe holds.
claude plugin validate .
The Finance and Legal crews are analytical tools. They produce models, reviews, and checklists to inform a decision — not licensed financial or legal advice. Every output states its assumptions and warrants professional review before anyone relies on it.
The Marketing crew will not build dark patterns. Fake countdowns, confirmshaming, obstructed cancellation, and pre-ticked consent are out of scope regardless of measured lift — they are illegal in a growing number of jurisdictions and they cost more than they earn. Ask for one and you get the legitimate version that achieves the same goal.
OmniSkill is MIT.
Ten bundled skills stay under their own licenses — MIT for superpowers, grill-me, and motion-ui; Apache-2.0 for the anthropics/skills and claude-plugins-public skills; SIL OFL for the 54 fonts in canvas-design. Every copy is unmodified and ships its license file. Full attribution: THIRD_PARTY_NOTICES.md.
xlsx and docx are deliberately not bundled — they are source-available and proprietary, so redistribution here would not be permitted.