ScrapingBee published a guide to using Agent Skills to make coding agents more reliable, and it contains the cleanest one-sentence separation of the two things everyone is currently conflating: “a skill teaches the agent how to approach a task, while an MCP server or other tool gives the agent an external capability.” Procedure versus capability. A skill is a decision framework written down; an MCP server is a door. Anyone building both should have that sentence pinned somewhere, because the design mistakes run in both directions — encoding a workflow as a tool call, or expecting a markdown file to reach a system it has no access to.
The mechanics are stated with useful precision. A skill is a SKILL.md — YAML frontmatter plus markdown instructions, with a 64-character ceiling on the name and 1,024 on the description — optionally carrying a requirements.txt and scripts, loaded by progressive disclosure so the deep material only arrives when it is needed. The advice is appropriately unglamorous: start with one to five skills rather than installing all 24 from a collection like addyosmani/agent-skills, keep the core workflow in SKILL.md and push depth into referenced files, and review a third-party skill the way you would review a dependency. Their honest closing position is the right one — skills “will not make coding agents perfectly reliable, but they can make good engineering habits more consistent.” And buried in the scraping advice is a sentence that belongs in every data-quality checklist: “a blank application shell or block page can still return valid HTML. Verify the fields you actually need.”
The catalog reads six API pages for ScrapingBee, and the parameters the post tells you to use sparingly are what those pages document — the HTML API where render_js and premium_proxy and wait_for live, the AI Web Scraping API behind ai_extract_rules, plus the Google Search API, the Data Extraction API and the Screenshot API. Turning on JavaScript rendering only when evidence demands it is cost advice on an API that meters exactly that.
ScrapingBee scores 42.5, developing on the Kin Score, with contract quality at 55.1 and discoverability at 68.5 against operational transparency at 10.5 and contract governance at 0.0. Agent Readiness is 23.2, agent-aware — and the two dimensions this post is about land on opposite sides. mcp_server is lit: they ship the door. agent_skills is unlit: they published the best available explanation of what a skill is and have not published one. That is the more interesting half of their own distinction, because a scraping API is an unusually good candidate for a skill. Every hard-won piece of judgment in that post — when render_js earns its cost, why a 200 with an empty shell is not a success, which fields to verify before trusting a row — is procedure, not capability. It is exactly what does not fit in a tool description.