The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes local file names and contents to determine project types and populate templates, which creates a surface for indirect prompt injection from potentially untrusted project data.
Ingestion points: Reads directory structure and file contents (e.g., package.json, pyproject.toml) during Step 2 and Step 3 to detect project markers.
Boundary markers: No explicit delimiters are mentioned for processing file content, but Step 5 provides a human-in-the-loop confirmation step before any file writes occur.
Capability inventory: The skill has file-writing capabilities across the local repository in Step 6.
Sanitization: The skill does not specify sanitization for ingested data before using it to select templates or fill variables.
[EXTERNAL_DOWNLOADS]: The generated CI templates include references to official GitHub Actions and established community tools.
Evidence: REFERENCE.md contains templates referencing actions/checkout@v4, actions/setup-node@v4, actions/setup-python@v5, actions/setup-go@v5, dtolnay/rust-toolchain@stable, and golangci/golangci-lint-action@v6. These are well-known and trusted services used for standard development workflows.
[COMMAND_EXECUTION]: The skill executes a local shell command to retrieve the user's identity for license attribution.
Evidence: Step 4a in SKILL.md specifies running git config user.name to populate the license holder field.

repo-scaffold