The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted project data (configuration files and source code) and external changelogs to generate its report, creating an injection surface.
Ingestion points: Reads repository files such as package.json, sanity.config.ts, and tsconfig.json (as detailed in detect.md), as well as source code patterns.
Boundary markers: SKILL.md defines explicit rules for trust, including "Read and report. Never modify," and requires separating verified findings from inferences.
Capability inventory: Utilizes shell commands like find, grep, cat, npm view, npm pack, and npx --no-install tsc for project discovery and data retrieval (found across detect.md and plugins.md).
Sanitization: The agent is instructed to verify export names, signatures, and version existence against live registry manifests rather than assuming safety from the input data.
[EXTERNAL_DOWNLOADS]: Fetches package metadata, manifests, and changelogs from the npm registry and Sanity's official GitHub repositories. These operations target well-known and established technology services.
[COMMAND_EXECUTION]: Executes several read-only shell commands to inspect the local development environment, project structure, and dependency tree.
[DYNAMIC_EXECUTION]: Uses node -e to programmatically evaluate semantic version ranges and npx --no-install tsc to resolve inherited configuration values from the project's TypeScript setup.

sanity-studio-upgrade