The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external outline.yaml and fields.yaml files and interpolates the content directly into prompts for sub-agents, creating a surface for instruction overrides.
Ingestion points: In Step 1, the skill reads an external outline.yaml file from the working directory. In Step 3, it reads a fields.yaml file. Data from these files, specifically the item_related_info and topic fields, is interpolated into the prompt template for background agents.
Boundary markers: The constructed prompt uses standard Markdown headers (e.g., ## Task, ## Field Definitions) as delimiters. However, it lacks explicit "ignore embedded instructions" warnings or advanced boundary markers to prevent the sub-agent from accidentally or maliciously obeying commands hidden within the research outline content.
Capability inventory: The skill is granted extensive capabilities including Bash for shell execution, Write for file persistence, and Task for launching additional autonomous agents.
Sanitization: While the output filename (item_name_slug) is sanitized to remove special characters, the descriptive fields and research data are interpolated into the prompt without validation or escaping.
[COMMAND_EXECUTION]: The skill instructs the agent to execute a Python validation script using the Bash tool.
Evidence: The prompt template in Step 3 requires the execution of python ~/.claude/skills/research/validate_json.py to verify the output. This pattern relies on a local script residing in the user's home directory.

research-deep