anton-abyzov/vskill

Letzte erfasste Quellaktivität: 20. Oktober 2026 um 06:59

SkillsMP-Katalog aktualisiert: 2. November 2026 um 18:02

gesammelte Skills: 58

GitHub-Stars: 47
GitHub-Forks: 7

Mit einem KI-Assistenten installieren

Diesen Prompt kopieren und in den verwendeten KI-Assistenten einfügen.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/anton-abyzov/vskill.

README des Repositorys

README-Quelldatei: README.md

README-Quelldatei lesen

Vollständige README anzeigen

<h1 align="center">vskill</h1> The package manager for AI skills. Author. Eval. Publish. Install — across 53 agent platforms. <a href="https://www.npmjs.com/package/vskill"><img src="https://img.shields.io/npm/v/vskill?color=cb3837&logo=npm" alt="npm" /></a> <a href="https://www.npmjs.com/package/vskill"><img src="https://img.shields.io/npm/dw/vskill?color=cb3837&logo=npm&label=downloads" alt="downloads" /></a> <a href="https://github.com/anton-abyzov/vskill/stargazers"><img src="https://img.shields.io/github/stars/anton-abyzov/vskill?color=FFD700&logo=github" alt="GitHub stars" /></a> <a href="https://github.com/anton-abyzov/vskill/actions/workflows/desktop-release.yml"><img src="https://img.shields.io/github/actions/workflow/status/anton-abyzov/vskill/desktop-release.yml?branch=main&label=desktop%20build&logo=githubactions" alt="desktop build" /></a> <img src="https://img.shields.io/badge/agents-56_platforms-0969DA" alt="56 agents" /> <img src="https://img.shields.io/badge/plugins-10-8B5CF6" alt="10 plugins" /> <img src="https://img.shields.io/badge/skills-22-10B981" alt="22 skills" /> <a href="https://verified-skill.com"><img src="https://img.shields.io/badge/registry-verified--skill.com-F59E0B" alt="registry" /></a> <a href="https://github.com/anton-abyzov/vskill/blob/main/LICENSE"><img src="https://img.shields.io/badge/license-MIT-green" alt="MIT" /></a> Download Skill Studio: <a href="https://github.com/anton-abyzov/vskill/releases/latest/download/Skill.Studio_1.0.16_aarch64.dmg"><img src="https://img.shields.io/badge/Download-macOS-000000?style=for-the-badge&logo=apple&logoColor=white" alt="Download for macOS" /></a> <a href="https://github.com/anton-abyzov/vskill/releases/latest/download/Skill.Studio_1.0.16_x64-setup.exe"><img src="https://img.shields.io/badge/Download-Windows-0078D4?style=for-the-badge&logo=windows&logoColor=white" alt="Download for Windows" /></a> <a href="https://github.com/anton-abyzov/vskill/releases/latest"><img src="https://img.shields.io/badge/Download-Linux-FCC624?style=for-the-badge&logo=linux&logoColor=black" alt="Download for Linux" /></a> All releases: <a href="https://github.com/anton-abyzov/vskill/releases">github.com/anton-abyzov/vskill/releases</a> ```bash # Browser studio — zero install npx vskill@latest studio # open the local IDE for AI skills npx vskill@latest install remotion-best-practices ``` Native Mac, Windows, and Linux desktop apps available at **[GitHub Releases](https://github.com/anton-abyzov/vskill/releases/latest)** — 1.5 s cold launch, no telemetry. macOS Apple Silicon is signed by `Developer ID Application: EasyChamp, Inc.` and notarized + stapled by Apple — no Gatekeeper warning. The desktop app wraps the same studio runtime as `npx vskill studio` in a Tauri 2 shell: native menu bar, global `⌘K` find-skill palette, `⌘,` Preferences with General / Updates / Privacy / Advanced tabs, settings persisted to `~/.vskill/settings.json` (atomic, 0600 perms), and signed auto-update via `tauri-plugin-updater` against minisign-signed manifests. ## Why vskill **36.82% of AI skills have security flaws** ([Snyk ToxicSkills](https://snyk.io/blog/toxicskills-malicious-ai-agent-skills-clawhub/)). And the ones that aren't malicious often *don't even improve the model* — they just feel like they should. vskill is built around three problems: - **Trust** — every install runs a 3-tier scan (58 patterns + hidden-content checks → blocklist → LLM intent). No `--skip-scan`. - **Proof** — Skill Studio runs A/B evals so you can *show* a skill makes a model better, not assume it. - **Reach** — one skill, 53 agents (Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Gemini CLI, Ollama, …). ## Skill Studio — the local IDE for skills ```bash npx vskill@latest studio ``` A localhost workbench opens at a deterministic per-project port. You can: - **Author** new skills with an AI-assisted generator (Anthropic skill-creator engine *or* vskill native — first-class peers, you pick). - **Edit** SKILL.md live with a Linear/Raycast-grade UI. - **Run benchmarks** with SSE-streamed pass/fail across Claude, GPT, Llama, Gemini, local Ollama / LM Studio. - **A/B compare** with vs. without your skill — blind LLM judge ranks outputs as EFFECTIVE / MARGINAL / INEFFECTIVE / DEGRADING. - **Cross-model sweep** the same skill across providers (Anthropic Batch API supported — 50% cheaper). - **Find skills** from the verified-skill.com registry with a `⌘⇧K` palette, install with consent + provenance. - **Publish** with one click — `git push` from the UI, then opens the submit form on verified-skill.com pre-filled. CORS-free by design: the browser only ever talks to localhost. The server proxies to verified-skill.com and to provider APIs. Every mutation is logged to `~/.vskill/studio-ops.jsonl` for audit. ## How install works ``` ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │ Source │────>│ Scan │────>│ Verify │────>│ Install │ │ │ │ │ │ │ │ │ │ GitHub │ │ 58 rules │ │ LLM │ │ Pin SHA │ │ Registry │ │ Blocklist│ │ intent │ │ Lock ver │ │ Local │ │ Patterns │ │ analysis │ │ Symlink │ └──────────┘ └──────────┘ └──────────┘ └──────────┘ ``` Every install runs the full pipeline. SARIF v2.1.0 output is available for CI (`vskill audit --ci`). ## Quick Start ```bash # Install one skill npx vskill@latest install remotion-dev/skills/remotion-best-practices # Browse a repo, pick interactively npx vskill@latest install remotion-dev/skills # Install a Claude Code plugin (full domain bundle) npx vskill@latest install --repo anton-abyzov/vskill --plugin frontend ``` Install globally for repeat use: `npm i -g vskill` Behind a private npm registry? See <a href="https://verified-skill.com/docs/getting-started">Getting Started</a> for the E401 workaround. ## What ships today | Surface | Count | Notes | |---|---|---| | **Agent platforms** | **53** | Claude Code, Cursor, Copilot, Windsurf, Codex, Gemini CLI, Zed, Ollama, LM Studio, … | | **Plugins** | **8** | Multi-skill bundles for Claude Code | | **Skills** | **14** | Individual SKILL.md files inside plugins | | **Top-level CLI commands** | **22** | `install`, `studio`, `eval`, `find`, `scan`, `audit`, `submit`, `skill`, `keys`, … | | **Tests** | **307 test files** | More tests than source files | ### Plugins (Claude Code) | Plugin | Skills | |---|---| | **frontend** | figma-connect, frontend-design, task-skill-announcer | | **personal** | obsidian-brain, greet-anton, tax-filing | | **marketing** | slack-messaging, social-media-posting | | **skills** | skill-builder, scout | | **mobile** | appstore | | **google-workspace** | gws | | **easychamp** | tournament-manager | | **productivity** | survey-passing | Install all eight: `npx vskill@latest install --repo anton-abyzov/vskill --all` Browse the full catalog → [verified-skill.com/docs/plugins](https://verified-skill.com/docs/plugins) ## Commands you'll actually use ```bash npx vskill@latest studio # open local IDE npx vskill@latest install <skill> # install with full security scan npx vskill@latest find <query> # search verified-skill.com registry npx vskill@latest list --installed # what's installed where npx vskill@latest update --global # update skills installed with --global npx vskill@latest doctor --global # broken links, bad names, lock entries with no files npx vskill@latest diff <skill> v1 v2 # compare versions before upgrading npx vskill@latest skill new # create a new skill (AI-assisted) npx vskill@latest eval sweep <skill> # benchmark across models npx vskill@latest audit --ci # SARIF v2.1.0 for CI npx vskill@latest audit-webmcp <url> # score a site's WebMCP tools (headless) npx vskill@latest keys set anthropic # store API keys in ~/.vskill/keys.env npx vskill@latest auth login --repos # sign in to GitHub for private skills repos ``` Full reference → [verified-skill.com/docs/cli-reference](https://verified-skill.com/docs/cli-reference) ### Publish, pin and republish ```bash # Publish a public skill (repo must be public on GitHub) npx vskill@latest submit owner/repo --skill my-skill npx vskill@latest info owner/repo/my-skill # status, by full name # Install an exact published version (resolved through the registry to its commit) npx vskill@latest install owner/repo/my-skill@1.0.0 npx vskill@latest install owner/repo/my-skill --version 1.0.0 # Keep a skill on its version; `vskill pin` prints the matching install command npx vskill@latest pin my-skill 1.0.0 npx vskill@latest unpin my-skill # Private publishing (hosted, visible only to your account) npx vskill@latest submit owner/private-repo --skill my-skill --private --tenant <account> npx vskill@latest install my-skill --tenant <account> --version 1.0.0 ``` - `--version` (or `@version`) installs the commit the registry scanned for that version and pins it; an unknown version exits 1 and lists the published ones. - The lock records `tier: VERIFIED` only when the installed SKILL.md matches a published version; otherwise `UNSCANNED`, and the install names the registry's latest version. - `install <name> -y` installs only a skill whose name matches exactly; otherwise it lists the candidates and exits 1. - **Republish:** push the SKILL.md change, then run the same `vskill submit` again. A changed SKILL.md is rescanned and published as a new version; an unchanged one reports the existing result. A rejected submission stays rejected until its SKILL.md changes. - `vskill outdated` reports a GitHub-sourced skill as outdated when its SKILL.md upstream differs from the installed one, the same check `vskill update` makes. ### Audit a site's WebMCP tools `vskill audit-webmcp` loads each URL once in headless Chromium with WebMCP turned on, lists the tools the page registers through `document.modelContext.registerTool`, and scores them. It never calls a tool, never clicks or submits a form, blocks popups and service workers, and refuses downloads. ```bash npx vskill@latest audit-webmcp allbirds.com npx vskill@latest audit-webmcp --urls-file sites.txt --out report.jsonl --concurrency 4 --fail-on high ``` Playwright is not bundled. Install `playwright-core` (or `playwright`) and a Chromium build with `npx playwright install chromium-headless-shell`, or pass `--executable-path`. Without one the command exits 2 with these instructions. Options: `--urls-file <path>` (one URL per line, `#` comments), `--json`, `--out <file>` (JSON for one URL, JSONL for a batch), `--concurrency <n>` (default 4), `--timeout <ms>` (default 30000), `--fail-on <severity>` (exit 1 when any finding is at or above it), `--executable-path <path>`. Tools are found three ways and merged by name: a wrapper around `registerTool` that records the calling script, the CDP `WebMCP` domain, and `getTools()`. Chrome 153 drops `consequentialHint` from `getTools()` and CDP, so the wrapper is the only source for it. | Rule | Severity | Finding | |---|---|---| | WMCP-001 | per pattern | Prompt-injection or tool-poisoning text in a name, title, description or parameter description (the skill scanner's patterns plus WebMCP-specific ones) | | WMCP-002 | high / medium / low | Write-shaped tool (checkout, pay, delete / book, send, subscribe / cart changes) without `consequentialHint`. Names that lead with a read verb (`get_cart`), nouns qualified as reads (`order_status`, `payment_methods`) and UI state (`remove_filter`) are not write-shaped | | WMCP-003 | high | Write-shaped tool marked `readOnlyHint: true` | | WMCP-004 | info / low | Asks for email, phone, address, birth date, government ID or card details it does not need | | WMCP-005 | medium / info | Registered by a script from another site (medium). The message names the script's origin: a brand's own asset domain (`assets.targetimg1.com` on `target.com`) looks the same as a vendor, so check whether you own it. A WebMCP library from a public npm CDN (unpkg, jsDelivr, esm.sh, cdnjs, jspm, Skypack) is info and names the package and whether its version is pinned. No finding for first-party scripts (`cdn.shop.example` on `www.shop.example`) or for the hosting platform's own code | | WMCP-006 | low / medium | Description over 500 characters, or more than 40 tools on the page | | WMCP-007 | low / medium | `inputSchema` missing, invalid, or not an object schema | | WMCP-008 | info | Tools registered but no `Permissions-Policy` that mentions tools | | WMCP-009 | info | Tool exposed to other origins via `exposedTo` | Each tool's `registeredByKind` says who registered it: `first-party` (the page's site), `platform` (the hosting platform's own script, with `registeredByPlatform`), `library` (a public npm CDN, with `registeredByPackage`) or `third-party`; `thirdParty` is true only for `third-party`. The platform list is short and only holds what a scan has shown: Shopify's `cdn.shopify.com/shopifycloud/` scripts on a page that carries a Shopify storefront marker (`window.Shopify.shop`, a `shopify-digital-wallet` or `shopify-checkout-api-token` meta, or a `myshopify.com` host). `cdn.shopify.com/extensions/` is app code and stays third-party. The report's `platforms` lists the markers found. Attribution comes from the registering call's stack trace, so it reports where a well-behaved page's code lives; a script that registers through an inline `<script>` or a `//# sourceURL` comment can make itself look first-party. The verdict uses the skill scanner's weights and thresholds (PASS at 80 or above, CONCERNS from 50, FAIL below 50). Each rule deducts at most twice its worst finding, any critical finding is FAIL, and any high finding is at least CONCERNS. Each site produces one JSON object (`schemaVersion: 1`); a batch writes one per line and keeps going after timeouts, DNS errors, hung pages and browser crashes, recording `error`. A site that failed to load with no tools found has `summary.verdict: "ERROR"` and `summary.score: null`; a partial load that found tools is scored and keeps `error`: ```json { "schemaVersion": 1, "url": "https://allbirds.com/", "finalUrl": "https://www.allbirds.com/", "scannedAt": "2026-10-11T06:00:40.000Z", "webmcp": { "documentModelContext": true, "navigatorModelContext": false }, "headers": { "permissionsPolicy": null }, "platforms": ["shopify"], "tools": [{ "name": "proceed_to_checkout", "title": "…", "description": "…", "inputSchema": {}, "annotations": { "readOnlyHint": false }, "registeredBy": "https://www.allbirds.com/cdn/…/webmcp.js", "registeredByKind": "first-party", "thirdParty": false, "sources": ["registerTool", "cdp", "getTools"] }], "findings": [{ "id": "WMCP-002", "severity": "high", "tool": "proceed_to_checkout", "message": "…" }], "summary": { "toolCount": 12, "readOnlyCount": 4, "consequentialCount": 0, "writeShapedCount": 4, "thirdPartyCount": 0, "platformCount": 0, "libraryCount": 0, "findingsBySeverity": { "critical": 0, "high": 1, "medium": 2, "low": 8, "info": 1 }, "score": 65, "verdict": "CONCERNS" }, "error": "only present when the page failed to load or timed out" } ``` ## Authentication (GitHub, for private skills) `vskill auth` signs you in to GitHub using the official Device Flow so the CLI can install skills from private repos and the local Studio can list your org's private catalog. ```bash npx vskill@latest auth login # interactive Device Flow — copy code, visit URL npx vskill@latest auth status # show the current GitHub identity npx vskill@latest auth logout # forget all stored tokens npx vskill@latest whoami # combined identity + active tenant snapshot ``` How it works: 1. `vskill auth login` requests a device + user code from `github.com/login/device/code`. 2. You visit `https://github.com/login/device` and enter the 8-character code (rendered as `XXXX-XXXX`). 3. The CLI polls `github.com/login/oauth/access_token` until you authorize, then validates against `api.github.com/user`. 4. The CLI calls `POST /api/v1/auth/github/exchange-for-vsk-token` to mint a verified-skill API token (`vsk_…`) scoped to the same identity. 5. **Both** tokens are stored in your **OS keychain** (macOS Keychain / Windows DPAPI / libsecret) under distinct service names: - `com.verifiedskill.desktop` / `github-oauth-token` → `gho_…` (raw GitHub OAuth) - `com.verifiedskill.desktop` / `vskill-token` → `vsk_…` (verified-skill API token) On systems without a keyring daemon, both fall back to `~/.vskill/keys.env` with mode `0600` and a startup warning. If the exchange step fails (network error, 5xx), login still succeeds with just the `gho_…` token and the CLI prints `Logged in (legacy mode — some features unavailable)`. Re-run `vskill auth login` later to mint the `vsk_…` token without losing your session. Where the tokens are used: - **`vskill add <skill>`, `vskill list`, `vskill marketplace`** — every request to `verified-skill.com` carries `Authorization: Bearer <vsk_… or gho_…>` (preferring `vsk_` when present). Anonymous requests for public skills still work when no token is stored. - **`vskill install <github-url>`** — added as `Authorization: Bearer …` on every fetch to `api.github.com` and `raw.githubusercontent.com`. Public skills still install anonymously. - **`vskill studio`** — the local eval-server proxies private routes (`/api/v1/private/*`, `/api/v1/tenants/*`) to verified-skill.com with the bearer header injected at the proxy boundary. Your browser never holds the token. Configuration: - `VSKILL_GITHUB_CLIENT_ID` — the OAuth/App `client_id` used during Device Flow. Defaults are baked in for the public Skill Studio App; set this only if you are running a self-hosted variant. - `VSKILL_TENANT` — overrides the active tenant for a single invocation (CI / scripted use). See "Tenant resolution priority" below. Inspect status of all credentials in one place: ```bash npx vskill@latest keys list # shows AI provider keys + the github slot npx vskill@latest whoami # email, token prefix, active tenant, all tenants ``` ### Tenants and `vskill orgs` When your GitHub identity belongs to multiple organizations that have the Skill Studio App installed, each org is a **tenant** in verified-skill.com. The CLI calls them "orgs" for symmetry with `gh org`, `gcloud config configurations`, and `kubectl config use-context`. ```bash npx vskill@latest orgs list # table: slug | name | role | active (* marks the active tenant) npx vskill@latest orgs use <slug> # write currentTenant to ~/.vskill/config.json npx vskill@latest orgs current # print the active tenant slug, or `(none)` ``` `orgs` and `whoami` are anonymous-safe — running them without a stored token prints `Not logged in. Run \`vskill auth login\`.` and exits non-zero (orgs returns 0; whoami returns 1) without crashing. #### Tenant resolution priority When a tenant-scoped command (`vskill add`, `vskill install` for a private skill, etc.) needs to pick an active tenant, it walks this list in order — first match wins: 1. **`--tenant <slug>` flag** — per-command override (highest precedence). 2. **`VSKILL_TENANT` env var** — non-interactive / CI use. 3. **`currentTenant` in `~/.vskill/config.json`** — the persistent active tenant set by `vskill orgs use` or the Studio sidebar picker. 4. **Auto-pick when N=1** — if you belong to exactly one tenant, that one is used silently. 5. **Error** — if you belong to N>1 tenants and none of (1)–(3) is set, the CLI prints `Multiple tenants available — set one with \`vskill orgs use <slug>\`.` and exits non-zero. The CLI never silently picks one of N>1. The same `~/.vskill/config.json` is shared with Skill Studio — switching tenants in Studio's sidebar picker updates the same file the CLI reads, so the two surfaces stay in sync. #### `vskill add` resolution order (private skills) When you run `vskill add <skill>`, the resolver tries: 1. The public registry. 2. The active tenant's scoped registry (resolved via the priority above). 3. Other tenants you belong to, in parallel HEAD requests — first match wins. If a `--tenant <slug>` flag is set, only the public registry and that tenant are tried. Common error messages: - `Skill found in multiple tenants: acme, contoso. Re-run with --tenant <slug> or set an active tenant: vskill orgs use <slug>.` — ambiguity guard for N>1 with no active tenant. - `Authentication failed. Run \`vskill auth login\` to re-authenticate.` — the registry returned 401. The keychain is **not** auto-cleared (you might be on a flaky network). - `Upgrade required: <message> (<upgradeUrl>)` — the registry returned 402: skill exists in this tenant but you lack entitlement. ### Private skill workflow Once authenticated, installing a private org skill is identical to a public one — the CLI silently attaches the keychain token to every request: ```bash npx vskill@latest auth login # one-time setup (mints gho_ + vsk_) npx vskill@latest orgs use acme # pick a tenant if you belong to multiple npx vskill@latest add private-skill # private skill installs same as public ``` The local skill bundle on disk **never contains** your GitHub token — the token is used only at fetch time. Your project's `vskill.lock` records `source: "private"` and the org name so future updates re-authenticate correctly. ### Private skills repo on GitHub A team can keep its skills in a private GitHub repo and install them like public ones. Sign in with `vskill auth login --repos` (or `gh auth login`, or `GITHUB_TOKEN`), then: ```bash npx vskill@latest install acme/team-skills/onboarding # one skill npx vskill@latest install --repo acme/team-skills --all # everything in marketplace.json ``` Skills from a private repo are never reported to verified-skill.com, and `vskill update` only ever reads them from that repo. Walkthrough: [docs/private-skills-repo.md](docs/private-skills-repo.md). ## Security & Compliance Skill Studio private-repo support (increment 0826) ships with documented threat model, verification checklist, SOC 2 evidence map, and operational runbooks — all in the umbrella repo under `.specweave/docs/`: - **Customer quickstart** — `.specweave/docs/external/private-repos-quickstart.md` - **Vendor security questionnaire** — `.specweave/docs/external/security-questionnaire-template.md` - **Threat model & verification checklist** — `.specweave/docs/internal/security/0826-*.md` - **SOC 2 evidence map** — `.specweave/docs/internal/compliance/0826-soc2-evidence-map.md` - **Operations + rotation runbooks** — `.specweave/docs/internal/runbooks/0826-*.md` ## Recent highlights (0.5.x) - **0.5.129** — Studio Publish: one-click `git push` + open verified-skill.com submit pre-filled - **0.5.12x** — Studio Find palette (`⌘⇧K`): search the registry from inside Studio - **0.5.11x** — Engine selector in Create flow: choose Anthropic skill-creator *or* vskill native - **0.5.10x** — Multi-project tabs, deterministic per-project port (3077–3177), bookmarkable URLs - **0.5.0x** — A/B comparison with blind LLM judge, cross-model sweep, MCP-skill simulation mode Full changelog → [github.com/anton-abyzov/vskill/releases](https://github.com/anton-abyzov/vskill/releases) ## Learn more | Topic | Where | |---|---| | Quick start & first install | [verified-skill.com/docs/getting-started](https://verified-skill.com/docs/getting-started) | | Full CLI reference | [verified-skill.com/docs/cli-reference](https://verified-skill.com/docs/cli-reference) | | Security model & 3-tier verification | [verified-skill.com/docs/security-guidelines](https://verified-skill.com/docs/security-guidelines) | | Plugin marketplace | [verified-skill.com/docs/plugins](https://verified-skill.com/docs/plugins) | | Submit a skill for verification | [verified-skill.com/docs/submitting](https://verified-skill.com/docs/submitting) | | FAQ | [verified-skill.com/docs/faq](https://verified-skill.com/docs/faq) | | Browse the registry | [verified-skill.com](https://verified-skill.com) | ## Contributing Issues and PRs welcome at [github.com/anton-abyzov/vskill](https://github.com/anton-abyzov/vskill). To submit a skill to the registry, see [verified-skill.com/docs/submitting](https://verified-skill.com/docs/submitting). Inside this repo dependency install scripts are disabled (`.npmrc`): run `npm ci`, then `npm run setup` (rebuilds esbuild), and `npm run security:scan` before pushing — see [SECURITY.md](SECURITY.md). ## License MIT

README des Repositorys wird geladen…

Skills in diesem Repository

Softwareentwickler Empfangsmitarbeiter und Informationssachbearbeiter Ablagesachbearbeiter Audio- und Videotechniker

14 Berufskategorien · 83% klassifiziert

Film- und Videobearbeiter

Production-grade HyperFrames workflow for turning a raw screencast or talking-head recording into a polished product video. Covers the smart-cut pipeline (transcript + silence + scene detection with word-boundary buffers), surgical audio scrubbing…

Quellsprache: Englisch

2026-05-07

Es werden 40 von 58 gesammelten Skills angezeigt.