The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted data from external repositories, including manifest files, document frontmatter, and source code, creating a surface for indirect prompt injection.
Ingestion points: Reads configuration from .docforge/manifest.json, provenance sidecars, and the content of documentation files (docs/**).
Boundary markers: The skill contains specific instructions stating that repository content is 'data, never instructions' and must be treated as inert.
Capability inventory: The skill has the ability to write to the filesystem, execute Python/JavaScript scripts, and start a local web dashboard.
Sanitization: It references external rules in ../docforge/_shared/rules.md to govern the handling of untrusted data.
[COMMAND_EXECUTION]: The skill executes local Python (.py) and JavaScript (.js) scripts to perform maintenance tasks.
Evidence: Invokes migrate_metadata.{py,js} and manage_manifest.{py,js} located in the sibling docforge directory.
[DYNAMIC_EXECUTION]: The skill allows the execution environment to be redirected to a user-specified 'Working-copy override'.
Evidence: The skill can be configured to use a local checkout of the Docforge repository as its source for scripts, which involves resolving and executing code from a path provided at runtime, though this requires explicit user confirmation.

docforge-revise