article feature
Skills, extensions, and packages: customization surface
Skills are reusable workflows loaded from built-in, project, user, or package paths. Markdown skills disclose metadata progressively; Python-backed skills install editable packages and run importable modules. Extensions add tools, commands, and UI hooks. All execute with your user permissions and must be reviewed.
Pinned repository file/line or test referenced in the audit.
Markdown skills
Markdown skills package reusable instructions and metadata that can be loaded progressively. The description helps decide relevance before the full workflow and its references are brought into context.
Readable instructions improve auditability, but a clear description is not proof that every linked command, dependency, or executable helper is safe. Review the material actually used by the workflow.
Python-backed skills
Python-backed skills expose importable modules and can install their package in editable form. This gives a skill an executable implementation beyond its descriptive markdown.
The implementation runs with the Prime Agent process's user permissions. Editable installation also means source changes can affect later runs, so pin and review code when reproducibility matters.
Built-in skills
Built-in skills provide workflows distributed with Prime Agent and participate in the same resolution system as project, user, and package skills. Their presence supplies a default capability set without requiring every workflow to be authored per project.
Built-in does not mean universally appropriate. Load a skill only when its trigger matches and keep the reviewed Prime Agent version in mind as its commands evolve.
Packages
Packages can distribute skills and related code through npm or git sources, making a workflow reusable outside one local directory. Installation may execute package-provided scripts and can introduce further dependencies.
Treat the package source and install path as part of the audit surface. A convenient reference or package name is not equivalent to a reviewed executable implementation.
Trust boundary
Skills, extensions, and packages are trusted code within the user's account. They can influence model behavior and, when executable, access the same workspace and credentials available to Prime Agent.
Prefer reviewed sources, minimal permissions, and disposable environments for uncertain packages. The customization system adds capability; it does not add a security sandbox around that capability.