What is skills?

PlanetScale Skills is a pack of instruction files that turns an AI coding agent into a reviewer and operator for PlanetScale, a hosted database service. Rather than being a plugin or piece of custom software, it is written as plain markdown "SKILL.md" files following an open convention that agents such as Claude Code and Cursor already know how to read. Point a compatible agent at a PlanetScale organization, and the skills guide it through auditing configuration, query behavior, safety settings, and backup posture, then, with the user's approval, carrying out fixes. The workflow has four stages. First the agent assesses a database or an entire organization in read-only mode, checking things like branch and schema-change workflow, backup and restore readiness, stored credentials, query performance data, and open recommendations. Second, it produces a single markdown report where every finding is backed by the exact command that produced it, along with a suggested fix and a way to undo it if something goes wrong. Third, once a user approves specific recommendations by their ID, the agent executes the changes through PlanetScale's own built-in safety features, such as development branches, pull requests, and deploy requests with a window to revert. Fourth, for ongoing maintenance, the skills support scheduled automation that continually sweeps for new recommendations and turns them into ready-to-review pull requests. Setup requires installing and authenticating the PlanetScale command-line tool, then copying the skill files into whichever directory the user's AI agent reads skills from, such as a Cursor or Claude Code folder. A setup script is provided that detects installed agents automatically and installs to all of them at once. The skills work best when the agent is run from inside the actual application repository that talks to the database, because then it can trace slow queries back to the specific code that causes them and fix that code directly, not just flag it in a report. By default, every action beyond the initial read-only assessment requires explicit user approval before anything changes. There is also an opt-in fully autonomous mode where a user can authorize the agent to apply an entire set of recommendations without approving each one individually, including in production, but only after typing an explicit risk acknowledgment naming the scope of what is allowed. For unattended scheduled runs, a similarly explicit authorization file must be committed to the repository ahead of time, specifying who owns the automation, what it is allowed to touch, and when that permission expires.