The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script (scripts/run.py) to automate data extraction. This is a standard implementation using argparse and pathlib for file operations.
[DATA_EXPOSURE_AND_EXFILTRATION]: No network access is requested or used. The skill only interacts with local project files such as outline/evidence_drafts.jsonl and citations/ref.bib, and writes to outline/anchor_sheet.jsonl. No sensitive system files or credentials are accessed.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from external files, presenting a potential surface for indirect injection.
Ingestion points: Evidence snippets and comparisons are read from outline/evidence_drafts.jsonl in scripts/run.py.
Boundary markers: The data is processed and stored in structured JSONL formats, which provides clear boundaries between different data objects.
Capability inventory: Capabilities are limited to filesystem reads and writes within the local workspace directory; the skill lacks network access or the ability to execute arbitrary commands.
Sanitization: The script filters content using specific regex patterns for quantitative data (e.g., searching for digits or benchmark-related keywords) and implements text truncation via the _trim function, which reduces the risk of propagating large or complex prompt injection payloads.

anchor-sheet