The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied research data (questions, source metadata, and tags) via external JSON files which are then processed by local Python scripts. The lack of explicit boundary markers or sanitization logic in the scripts creates a vulnerability surface for indirect prompt injection if the source data contains malicious instructions designed to influence the agent's synthesis.
Ingestion points: question.json, sources.json, and tagged_sources.json are used as inputs for the search strategy, quality scoring, and synthesis workflows.
Boundary markers: The instructions do not define delimiters or specific warnings to ignore instructions embedded within the processed research content.
Capability inventory: The skill executes local Python scripts (search_strategy_builder.py, source_quality_scorer.py, thematic_synthesis_builder.py) to generate markdown reports from the ingested data.
Sanitization: The Python scripts perform structural parsing but do not sanitize or filter the content of the string fields before they are output as markdown and read back into the agent context.

litreview