Claude Skills Repo
l3ad3r1/Claude-skills-repo
Skills reutilizáveis para automação, manipulação de docs e acesso à web
Ver no GitHubColeção open-source de Agent Skills reutilizáveis para Claude, incluindo ferramentas para conversão de documentos, manipulação de PDFs/Word/Excel/PowerPoint, automação com navegador, construção de MCP servers e acesso à internet. Cada skill é um pacote de instruções que ensina ao Claude como executar tarefas especializadas de forma confiável.
Instalação
npx convez-ai l3ad3r1/Claude-skills-repo1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Converter documentos variados (PDF, Word, Excel, PowerPoint) para Markdown para pipelines de análise
- Automatizar navegação web e interação com formulários via browser controlado
- Manipular e processar arquivos Office (criar, ler, editar documentos)
- Construir MCP servers e Agent Skills customizadas
- Clonar websites inteiros em codebases Next.js + Tailwind
- Dar ao agente acesso a internet em tempo real (web, redes sociais, feeds)
Veredito

Repositório maduro e bem-estruturado que oferece skills prontas para produção cobrindo automação de documentos, web scraping, construção de ferramentas e acesso à internet. A qualidade de implementação é alta, com dependências de bibliotecas consolidadas; ideal para quem quer estender capacidades do Claude com ferramentas reutilizáveis.
README
A tradução para o português deste README ainda está na fila.
Claude Skills
A collection of open-source Agent Skills for Claude — reusable instruction packages that teach Claude how to perform specialized tasks reliably.
Skills in this repo
| Skill | What it does |
|---|---|
markitdown-converter | Converts files of almost any format (PDF, Word, PowerPoint, Excel, HTML, CSV/JSON/XML, images, audio, EPub, ZIP archives) into clean, structure-preserving Markdown for LLM and text-analysis pipelines — then optionally analyzes the result. |
dev-browser | Drives a real browser (navigate, click, fill forms, screenshot, full Playwright API) via sandboxed JavaScript, wrapping the open-source dev-browser CLI by Sawyer Hood (MIT). |
pdf-toolkit | Merge, split, rotate, watermark, extract text/tables, and OCR PDFs — built on pypdf, pdfplumber, and OCRmyPDF. |
docx-toolkit | Create, read, and edit Word .docx documents (headings, tables, images, runs) — built on python-docx. |
xlsx-toolkit | Create, read, and edit Excel .xlsx workbooks (sheets, formulas, formatting, charts) — built on openpyxl. |
pptx-toolkit | Create, read, and edit PowerPoint .pptx decks (slides, bullets, images, notes) — built on python-pptx. |
mcp-server-builder | Build MCP servers (tools/resources/prompts) in Python (FastMCP) or Node/TS — with a working template and best practices. |
skill-builder | Author, validate, and package Agent Skills into installable .skill archives — with bundled validator and packager. |
animejs | Build web animations with anime.js v4 (MIT, by Julian Garnier) — animate DOM/CSS/SVG/JS objects, timelines, staggers, scroll triggers, draggables, springs, and text; includes a v4 API cheat-sheet and a runnable CDN demo. |
clone-website | Reverse-engineer and clone any website into a Next.js + shadcn/ui + Tailwind codebase — extracts assets/CSS/content section-by-section via a browser MCP and dispatches parallel builder agents in worktrees. Adapts the ai-website-cloner-template by JCodesMore (MIT); needs a browser MCP and the upstream Next.js scaffold. |
agent-reach | Give the agent live internet access — installs and drives the open-source Agent Reach CLI to read/search web pages, Twitter/X, YouTube, Reddit, GitHub, RSS, Bilibili, Xiaohongshu, LinkedIn, podcasts, and Exa web search through one zero-API-fee tool with automatic backend failover. By Panniantong (MIT); the CLI is installed from upstream, not vendored. |
markitdown-converter
Convert almost any document into clean Markdown, then either get the file back or have Claude analyze its contents — depending on what you ask for.
Supported formats
PDF · Word (.docx) · PowerPoint (.pptx) · Excel (.xlsx, .xls) ·
HTML · CSV / TSV · JSON · XML · plain text · images (EXIF + OCR) ·
audio (EXIF + transcription) · EPub · ZIP archives (iterates over contents).
Batch inputs (a folder or a ZIP of mixed files) are merged into one combined
Markdown file, with each document introduced by a # <filename> header and
separated by --- rules.
Install
Claude.ai (and Claude apps). Download markitdown-converter.skill from the
Releases page, then upload it under
Settings → Capabilities → Skills.
Claude Code. Copy the skill folder into your project (or personal) skills directory:
# project-level
mkdir -p .claude/skills
cp -r skills/markitdown-converter .claude/skills/
# or personal, available everywhere
cp -r skills/markitdown-converter ~/.claude/skills/
Then ask Claude to convert or markdownify a file and it will trigger automatically.
Standalone CLI usage
The bundled script works on its own, no Claude required:
pip install "markitdown[all]"
# single file -> single .md
python skills/markitdown-converter/scripts/convert.py report.pdf -o report.md
# folder or ZIP of mixed files -> one combined .md
python skills/markitdown-converter/scripts/convert.py mixed_docs.zip -o combined.md
python skills/markitdown-converter/scripts/convert.py ./docs -o combined.md
The script handles single files, multiple files, folders, and ZIP archives, and keeps converting the rest of a batch even if one file fails.
Known limits
- Scanned / image-only PDFs may come out empty — MarkItDown reads the embedded text layer, it is not a full PDF OCR engine.
- Plain PDFs carry no heading metadata, so a visual title may convert to
plain text rather than a
#heading. - Audio transcription and YouTube URLs need network access and optional dependencies.
- Complex tables (merged / multi-row headers) may render imperfectly; the script auto-repairs the common DOCX empty-header-row case.
dev-browser
Drive a real browser from the agent — navigate, click, fill forms, run the full
Playwright API, and take screenshots via sandboxed JavaScript. This skill
wraps the open-source dev-browser
CLI; the upstream tool is not vendored here — it's installed from npm.
Install
npm install -g dev-browser
dev-browser install # installs Playwright + Chromium
Use
dev-browser --headless <<'EOF'
const page = await browser.getPage("main");
await page.goto("https://example.com", { waitUntil: "domcontentloaded" });
console.log(await page.title());
EOF
Run dev-browser --help for the full LLM usage guide and API reference. See the
skill for Windows/PowerShell usage and --connect mode.
animejs
Write web animations with anime.js v4 — the modular,
ESM-first rewrite. The skill teaches Claude the v4 API (animate,
createTimeline, createDraggable, onScroll, svg, text, stagger,
utils, springs/eases) and the v3→v4 migration deltas.
Install
Claude.ai (and Claude apps). Download animejs.skill from the
Releases page, then upload it under
Settings → Capabilities → Skills.
Claude Code.
mkdir -p .claude/skills
cp -r skills/animejs .claude/skills/ # or ~/.claude/skills/ for personal use
What's inside
SKILL.md— when-to-use, install/import (npm + CDN), the module map, common recipes, and migration notes.reference.md— a per-module cheat-sheet with copy-pasteable snippets.examples/index.html— a self-contained, buildless demo that loads anime.js from a CDN.
The library itself is installed from npm (npm install animejs) or a CDN — it is
not vendored in this repo.
clone-website
Reverse-engineer and rebuild any website as a pixel-perfect Next.js + shadcn/ui + Tailwind v4 clone. The skill drives a browser MCP to extract design tokens, assets, computed CSS, and interaction models section-by-section, writes an auditable spec per component, and dispatches parallel builder agents in git worktrees.
Requirements
- A browser MCP (Chrome MCP, Playwright MCP, Browserbase, etc.).
- The Next.js scaffold from the upstream ai-website-cloner-template — create a repo from that template, then run the skill inside it.
Install
Claude.ai (and Claude apps). Download clone-website.skill from the
Releases page, then upload it under
Settings → Capabilities → Skills.
Claude Code.
mkdir -p .claude/skills
cp -r skills/clone-website .claude/skills/ # or ~/.claude/skills/ for personal use
Then run /clone-website <url1> [<url2> ...] from inside a project created from
the upstream template.
The Next.js scaffold and browser tooling are not vendored in this repo.
agent-reach
Give the agent eyes on the live internet. This skill installs and drives the
open-source Agent Reach CLI, which
wraps a fleet of free / open-source backends (Jina Reader, twitter-cli, yt-dlp,
OpenCLI, rdt-cli, bili-cli, gh, feedparser, Exa via MCP) behind one tool with
primary + fallback failover — so the agent can read and search content a
plain HTTP fetch can't reach.
Reaches
Web pages · YouTube transcripts · RSS/Atom · GitHub · global search (Exa) · Bilibili · V2EX · Xueqiu · Twitter/X · Reddit · Xiaohongshu · LinkedIn · Xiaoyuzhou podcasts. No-login channels work out of the box; login channels (Twitter/X, Reddit, Xiaohongshu, LinkedIn, podcasts) require credentials — use a dedicated secondary account, never your primary, since automated reading can get accounts banned.
Install
Claude.ai (and Claude apps). Download agent-reach.skill from the
Releases page, then upload it under
Settings → Capabilities → Skills.
Claude Code.
mkdir -p .claude/skills
cp -r skills/agent-reach .claude/skills/ # or ~/.claude/skills/ for personal use
The skill then installs the upstream CLI on demand:
pipx install https://github.com/Panniantong/agent-reach/archive/main.zip
agent-reach install --env=auto
agent-reach doctor # check which channels are live
Credentials and tool repos live under ~/.agent-reach/ (config at
~/.agent-reach/config.yaml, 600 perms, never transmitted). The Agent Reach
CLI and its backends target Unix-like environments and are not vendored in
this repo — they're installed from upstream (use WSL on Windows).
Credits
Skills packaged by Rinu (l3ad3r1) in collaboration with Claude (Anthropic). Each skill credits its upstream authors:
markitdown-converteris powered by Microsoft MarkItDown (MIT) — all credit for the underlying format support belongs to its authors.dev-browserwraps the dev-browser CLI by Sawyer Hood (MIT), brought to you by Do Browser.pdf-toolkitis built on pypdf (BSD), pdfplumber (MIT), and OCRmyPDF (MPL-2.0).docx-toolkit,xlsx-toolkit,pptx-toolkitare built on python-docx, openpyxl, and python-pptx (all MIT).mcp-server-builderuses the official MCP SDKs (MIT).skill-builderis an original work based on the public Agent Skills spec.animejsdocuments and wraps anime.js v4 by Julian Garnier (MIT) — all credit for the animation engine belongs to its author and contributors; the library is installed from npm/CDN, not vendored.clone-websiteadapts the ai-website-cloner-template by JCodesMore (MIT) — the SKILL.md instructions are reproduced from the upstream template; the Next.js scaffold and browser tooling are not vendored.agent-reachdocuments and drives Agent Reach by Panniantong (MIT) — all credit for the CLI and its backend integrations belongs to its author and contributors; the CLI is installed from upstream, not vendored.- These skills are original, permissively-licensed implementations — they do not include or derive from any proprietary skill content.
- Built for the Agent Skills format.
Maintenance
This repo is the canonical source of truth for skill content. The companion Hermes-skills repo repackages the same skills into category folders with richer frontmatter, so it is a derived copy: every skill added or changed here should be mirrored there.
Before cutting a release, verify the two repos are in sync:
python tools/check_parity.py # assumes Hermes checked out at ../_hermes
python tools/check_parity.py ../Hermes-skills # or pass the path explicitly
The check errors if a skill or bundled file exists in one repo but not the
other (the kind of drift that is easy to miss), and warns when a SKILL.md
body differs — some skills intentionally carry Hermes-specific wording.
License
The skill packaging in this repo is MIT © 2026 l3ad3r1. Each wrapped or adapted upstream project (MarkItDown, dev-browser, anime.js, ai-website-cloner-template, Agent Reach) remains under its own MIT license held by its respective authors.