Developer's custom Claude Code SEO Skill fails on first-day tests

On 2026-11-01 at 13:00, the AI section of DEV Community published a first-hand account from a developer whose self-built SEO Skill for Claude Code failed in its very first day of testing, with two bugs surfacing on pages he had considered safe.

The developer explains how the setup works. Claude Code skills, he writes, are simply folders: a SKILL.md file containing a name and a description, plus any scripts you want the model to run. Claude loads the instructions and follows them only when the description matches what the user asked for. That, he argues, is exactly why SEO suits this mechanism. Most on-page checks are deterministic, and a model is far better at explaining a result than at counting characters inside a <title> tag. So he built one: a script performs deterministic checks against a single URL, and the JSON report is handed to Claude, which turns it into a fix list.

For verification he used five real pages and one fake page. On day one the Skill still got things wrong on two of the pages he had assumed were safe. He also notes that the Skill sits in the project or user directory, and is loaded only when its description matches the task at hand.

The report does not specify what caused the two bugs, which pages or checks were affected, whether a fix followed, or whether the Skill has been shared publicly. The available information therefore stops at the point where the developer found two errors on day one.