The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from outline/evidence_drafts.jsonl to create anchors that dictate the content of a downstream writing task. \n- Ingestion points: Untrusted content is read from outline/evidence_drafts.jsonl by scripts/run.py. \n- Boundary markers: No explicit data boundary markers are used in the output file, but the usage policy defines strict constraints. \n- Capability inventory: The skill script executes file read and write operations within the workspace. \n- Sanitization: The script implements _META_ANCHOR_RE to identify and discard instructional phrases or prompt-like patterns found in the evidence data, reducing the risk of indirect injection.

anchor-sheet