The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: Several tools in the collection are designed to ingest and process untrusted external data, which is then presented to the AI agent. This creates a surface for instructions embedded in analyzed data to override agent behavior.
Ingestion points: code-to-prd/scripts/codebase_analyzer.py reads entire codebase directories; product-manager-toolkit/scripts/customer_interview_analyzer.py reads interview transcripts; research-summarizer/scripts/extract_citations.py reads research documents.
Boundary markers: The skills do not implement specific delimiters or 'ignore' instructions when interpolating external content into the agent context.
Capability inventory: The skill set includes file system write capabilities (saas-scaffolder/scripts/project_bootstrapper.py) and local command execution (roadmap-communicator/scripts/changelog_generator.py).
Sanitization: Untrusted content is parsed and summarized without filtering for potential prompt injection markers.
[COMMAND_EXECUTION]: The roadmap-communicator/scripts/changelog_generator.py script executes local commands to interact with the repository's history.
Evidence: Uses subprocess.run(['git', 'log', '--pretty=format:%s', commit_range]) to fetch commit messages.
Mitigation: The script uses a list-based argument format and does not enable shell=True, which prevents basic shell injection attacks.

product-skills