The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves ingesting and analyzing untrusted content from the repository it is run against, such as README files, architecture docs, and source code.
Ingestion points: In Phase 0a, it specifically searches for and reads files like README.md, ARCHITECTURE.md, and .cursorrules. In Phase 2, it performs a deep scan of the entire codebase.
Boundary markers: The skill implements a mandatory "Phase 1" gate where it must stop and present an inventory to the user for confirmation before generating any documentation. It also emphasizes a grounding discipline where every claim must cite a specific path:line, facilitating user verification.
Capability inventory: The skill has the capability to write multiple files to the local filesystem (under docs/system-flow/) and execute local tools including git, python3, and various infrastructure CLIs (kubectl, pulumi, flux).
Sanitization: While the skill does not explicitly describe input sanitization for LLM context, its design forces the agent to treat documentation as secondary to the actual source code and requires explicit citations for all findings.
[COMMAND_EXECUTION]: The skill uses several local command-line operations to perform analysis and build deliverables.
It executes bundled Python utility scripts (get_docs_dir.py and build_html.py) to manage paths and assemble the final HTML documents.
It uses git diff to support incremental updates by identifying changed files since the last mapping run.
It includes optional playbooks for infrastructure tools (kubectl, pulumi, flux) that involve executing read-only introspection commands. These are strictly gated by explicit user opt-in and a requirement to confirm the target environment (e.g., kube context) before execution.

system-flow-mapper