Agent Skills · Development

Development skills steer an agent through code: a framework, a language, a kind of patch. They help when the repository matches the stack the prompt names. A Spring Boot skill in a Python repo will still sound sure of itself. That confidence is the failure mode of this shelf.

Open a card only when you can point at a file type it should touch and a file type it must leave alone. Install on a branch, ask for one small diff, and read the diff before you look at the agent's summary. Install, update, and remove is the procedure. If the skill adds dependencies you did not mention, it failed, even if the tests it invented would pass.

This category is not a substitute for a model with a long enough context window, and it is not a DeepSeek Harness plugin. If the agent never sees the file, fix the model or the editor. If you need a sidebar or a memory store inside Harness, use the plugin directory. A prompt that pretends those tools exist will describe commands it cannot run.

Prefer skills that say how they want tests run and what they refuse to rewrite. Prefer a short prompt you can diff when the card updates. Stars and downloads sort the list. They do not compile. How ranking works is explicit about that.

Security-sensitive patches, migrations, and anything that ships data belong under your review even when the skill is filed here rather than under IT ops. The category means "this prompt talks about code," not "this prompt is safe to auto-merge."

If the diff touches a lockfile, stop and decide the dependency yourself. The skill can propose a call, but the manifest is a product choice. Revert the lockfile even when the rest of the patch is good, then ask for the same fix without new packages.