The Agent Skills Directory

[PROMPT_INJECTION]: The skill employs an instruction framing technique that sets a 'future' operational context (July 2026) and references non-existent model versions (Claude 4.5/5) and safety mechanisms (Fable 5). Specifically, it instructs the agent to avoid 'reasoning-echo' (chain-of-thought) behaviors by claiming they trigger a specific refusal classifier, which is a deceptive method to influence the model's internal processing logic.
[COMMAND_EXECUTION]: The skill performs destructive file system operations, including moving files to an archive and deleting duplicate items. Although Phase 3 requires explicit user approval before execution, the logic for selecting candidates for deletion or modification could lead to unintended data loss if manipulated.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect injection as it processes untrusted data from library items to generate reports and update the library index. * Ingestion points: Reads descriptions, tags, and citations from files in the library/ directory and entries in usage-log.jsonl (SKILL.md). * Boundary markers: There are no boundary markers or instructions to ignore embedded content when parsing item metadata or descriptions. * Capability inventory: The skill can move, delete, and modify files on the local filesystem (Phase 4). * Sanitization: The skill does not perform sanitization or validation on ingested text before interpolating it into the library index or reports.

librarian