miro.medium.com

Claude Code Skills: One Markdown File, 50 Listings, Zero Code

Claude Code Skills, SKILL.md, web scraping, AI workflow, structured data extractionDev

Anthropic 공식 — Agent Skills 개요

Firecrawl — Claude Code Skill 작성 튜토리얼

Santiago Valdarrama — "Claude Code로 라이브 웹 데이터를 테이블로 받는다"

Here's what Santiago Valdarrama did with Claude Code — "Pull 50 two-bedroom listings from 5 sites at once and get a clean, organized spreadsheet back." Zero lines of code written.

The trick is a new format called Skills. Drop a single markdown file into Claude Code and "scrape live data from the web" becomes an automatic capability. Claude, ChatGPT, and Gemini all adopted this pattern around spring 2026 — but the real value shows up when you use it as a coding tool.

What Is It?

A Claude Code Skill is a single file with YAML front matter + a markdown body. It lives at ~/.claude/skills/{name}/SKILL.md. Claude reads the description field, listens to what you're asking, and automatically picks which Skill to use. You don't have to type "/web-scrape" or any slash command — context triggers it.

Santiago's LinkedIn post got attention for a simple reason. "You can use Claude Code to pull live, structured data from any website... organized in tables you can work with immediately. Not a wall of text." His example: two-bedroom listings in a specific area, posted within 48 hours, with price and square footage extracted, normalized across multiple platforms — all from a single natural-language request.

Here's the thing — this is where it diverges from IDE-based tools like Cursor or Windsurf. A Skill is an abstraction unit for tool calls. Write the SKILL.md once and that capability is automatically available across every project. No more pasting the same instructions into every prompt.

SKILL.md structure
YAML front matter (name, description, allowed-tools) + a markdown body that describes the workflow.
The description field is everything
Claude uses this one sentence to decide whether to trigger the Skill. Be specific — something like "Use when user asks for live web data extraction."
Declare your tools
List external APIs like Firecrawl, Nimble, or Zyte in allowed-tools. Auth keys go in environment variables.
Specify the output format
Lock in deterministic instructions in the body — like "always return a markdown table."

What Changes?

Approach	Cursor / Standard IDE	Claude Code Skill
Capability persistence	Rewrite the prompt for each project	Write SKILL.md once → works everywhere
Trigger	Slash commands or explicit calls	Automatic context matching
Output	Markdown text	Structured data (JSON, CSV, tables)
External API	Inject keys on every call	Declared once in SKILL.md
Sharing	Tied to a codebase	Deploy to the team as a single file

Zyte shipped a Claude Skill using the same syntax — HTML fetching → AI parsing → selector generation → structured extraction, all four steps packed into a single Skill. Firecrawl published an official tutorial on writing Skills on their blog. Nimble earned Santiago's "10/10" endorsement as a recommended tool.

The key insight isn't "scraping automation" — it's "packaging capability." The same pattern works for anything repetitive: DB queries, PDF parsing, API integrations, meeting notes. Any repeated workflow can be extracted into a SKILL.md.

💡 Skill Directory Structure Tip
~/.claude/skills/web-scraper/SKILL.md is the main file, but you can add subdirectories like examples/, schemas/, and templates/ in the same folder — Claude will reference them automatically when needed. Think of SKILL.md as a table of contents, and keep the actual content separate to keep context light. That's the standard approach.

Getting Started

Check your Claude Code version
Skills require Claude Code 1.0+. Run claude --version to confirm.
Create the Skill directory
mkdir -p ~/.claude/skills/web-scraper. Use kebab-case for the name.
Write your SKILL.md
YAML front matter with name, description, and allowed-tools (curl, Firecrawl, etc.). Describe the workflow steps in the body.
Register external tools
Put API keys for Firecrawl, Nimble, etc. in .env and reference them from SKILL.md. FIRECRAWL_API_KEY as an environment variable is the standard approach.
Call it with plain English
Something like "Get all 2-bedroom apartments posted in the last 48h in [area]." Claude auto-triggers the Skill.

FAQ

(Auto-rendered by FAQSection — not included in content)

Deep Dive Resources

Firecrawl — Claude Code Skill Writing Tutorial A complete guide from SKILL.md structure to multi-function Skills firecrawl.dev

Santiago (@svpino) — Original LinkedIn Post "10/10 for agents needing web data" — the primary source for the rental listing example linkedin.com/svpino

Zyte — Supercharging Web Scraping with Claude Skills A 4-step Skill pattern: HTML → AI parsing → selector → structured data extraction zyte.com

yfe404/web-scraper An open-source scraping Skill with a 6-step quality gate — curl first, browser only when you need it github.com/yfe404/web-scraper

FAQ

Is Skills free to use?

The Skill itself is free — it's just a markdown file. What costs money is (1) your Claude Code subscription (Pro at $20 or Max at $100+) and (2) any external APIs the Skill calls. Firecrawl has a free tier, Nimble is in beta, and Zyte is usage-based. For a typical PM or marketer use pattern, Pro + Firecrawl free tier is a perfectly fine starting point. Heavy scraping will require Nimble or Zyte paid plans.

Can non-technical users write their own SKILL.md?

You can. The fastest approach: show Claude an existing SKILL.md example and say 'turn this into a Skill for my workflow.' There are solid examples in the yfe404/web-scraper and ComposioHQ/awesome-claude-skills repos. That said, you'll need some technical understanding for auth keys, environment variables, and rate limit handling. For your first Skill, start with something simple that doesn't require external APIs — that's the safe entry point.

Is it compatible with ChatGPT Skills or Gemini Agent Skills?

The file format is similar but not 100% interchangeable. Anthropic, OpenAI, and Google each tweaked their format slightly. That said, the core pattern — description-based auto-trigger + tool declarations + workflow body — is the same across all three, so migrating a Claude Skill to ChatGPT format takes about 30 minutes. Standardization is underway, so full compatibility looks likely down the road.

Is this legally safe? Won't it violate site Terms of Service?

It depends on the site, and it's more about how you use it than the Skill itself. Reading public pages at normal speed is generally fine. But (1) bypassing login-protected pages by force, (2) ignoring robots.txt, and (3) exceeding rate limits can violate ToS. For commercial use or redistribution, it's safer to go through legitimate scraping infrastructure like Firecrawl, Zyte, or Nimble — they handle ToS negotiation, authentication, and block avoidance on your behalf.

Can I deploy a Skill to my whole team?

Yes. Two approaches: (1) commit the entire ~/.claude/skills/{name}/ directory to a git repo for the team to clone, or (2) publish it to a marketplace like MCP Market or Composio. Never put auth keys or environment variables directly in SKILL.md — always split them into .env. If company security policy rules out external marketplace publishing, running it from an internal git repo is the standard approach.

Written by Rush

Tracking where business meets AI.

Did you find this reference helpful?

Get curated references delivered to your inbox weekly

Share this reference

Antioch — Meet the Cursor for Robot AI

Physical AI startups no longer need to rent warehouses or build million-dollar test facilities. Antioch brings software-speed development to robotics through cloud simulation — and just raised $8.5M seed to prove it.

Explore more AI workflow guides on similar topics

$20K and 12 AI Tools Built a $1.8B Telehealth Company — And Then the Red Flags Arrived

morningbrew.com

Medvi telehealth, AI startup leverage, GLP-1 startup, one-person unicorn, AI operations

$20K and 12 AI Tools Built a $1.8B Telehealth Company — And Then the Red Flags Arrived

Matthew Gallagher built Medvi, a GLP-1 telehealth startup, in 14 months with $20,000 and AI tools. 2 employees. 16.2% net margin. $401M in year one. Here's how the model works — and where it's breaking.

AI That Works While You Sleep — Automating Recurring Tasks with Claude Code Scheduled Task

substackcdn.com

What if your code review was already done when you woke up, and your newsletter

AI That Works While You Sleep — Automating Recurring Tasks with Claude Code Scheduled Task

What if your code review was already done when you woke up, and your newsletter sources were already organized? Here's how to automate recurring tasks with Claude Code Scheduled Task.

Next →Antioch — Meet the Cursor for Robot AI

Claude Code Skills: One Markdown File, 50 Listings, Zero Code

What Is It?

What Changes?

Getting Started

FAQ

Deep Dive Resources

FAQ

Share this reference

Antioch — Meet the Cursor for Robot AI

You might also like

$20K and 12 AI Tools Built a $1.8B Telehealth Company — And Then the Red Flags Arrived

AI That Works While You Sleep — Automating Recurring Tasks with Claude Code Scheduled Task