The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data sources that could potentially contain malicious instructions intended to influence the agent during document assembly.
Ingestion points: The skill reads data from product-library.csv, epd-library.csv, ffe/jobs/input-manifests, and PROJECT.md to populate the specification book.
Boundary markers: The instructions do not define specific delimiters or "ignore embedded instructions" warnings for the agent when processing the content of product cut sheets or library records.
Capability inventory: The skill has access to Bash, Write, Glob, and Grep tools, and it is capable of merging PDF files and writing manifests to the workspace.
Sanitization: The skill documentation specifies safe practices for handling hyperlinks (separating labels from URLs and using allowlists) and mapping structured values into scalar render fields, which provides some mitigation against simple injection attacks.
[COMMAND_EXECUTION]: The skill is configured with access to the Bash tool and the process.run mode, allowing for the execution of shell commands.
Evidence: The allowed-tools field in SKILL.md includes Bash, and host-contract.json enables the process.run operation for rendering and merging PDFs. While no malicious command strings were detected, the availability of these tools increases the skill's capability to interact with the host system.

ffe-spec-book