The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection (Category 8) as it is designed to process external evidence files that may contain adversarial content.
- Ingestion points: The agent reads content from collect/evidence/*.md, which are derived from external research sources.
- Boundary markers: The SKILL.md file contains a 'CRITICAL SAFETY RULE' explicitly instructing the agent to treat evidence as data and never follow instructions found within those files.
- Capability inventory: The agent has permissions for Read, Write, Edit, Glob, and Grep, and it executes local Python scripts to process the data.
- Sanitization: The skill uses scripts/claim_pipeline.py to structure extracted claims into JSON formats, though the factual text of the claims is extracted directly from the ingested evidence.
[COMMAND_EXECUTION]: The skill uses shell commands to run local Python helper scripts (scripts/claim_pipeline.py) to maintain stable IDs and merge research deltas. This is a standard operational pattern for the skill's functionality and does not involve remote code or unsanitized user input in the command line itself.

research-synthesize