The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data from scientific papers (PDF/HTML) and metadata from public genomic archives (SRA, ENA, GEO). This creates a vulnerability surface for indirect prompt injection attacks.
Ingestion points: Paper content extracted to 01_plan/paper_markdown/ and metadata records stored in 04_data/data_manifest.md.
Boundary markers: While the skill uses structured logging, it lacks explicit boundary markers or instructions to ignore embedded directives when processing paper content.
Capability inventory: The skill has the capability to write files, execute shell commands through the async_submit.sh utility, and manage containerized environments.
Sanitization: The async_submit.sh script utilizes printf %q to escape input variables when generating temporary shell scripts.
[COMMAND_EXECUTION]: The skill relies on the async_submit.sh utility to execute bioinformatics pipelines and tools as background processes. This mechanism is core to the skill's function for long-running analyses but grants the agent direct shell execution capabilities.
[EXTERNAL_DOWNLOADS]: The skill fetches datasets, paper supplemental files, and software dependencies from external scientific repositories (e.g., Zenodo, Figshare, bioRxiv) and container registries. The instructions prioritize user approval for these operations.
[DYNAMIC_EXECUTION]: The async_submit.sh script dynamically generates and executes shell scripts in the .task_status/ directory at runtime to manage task state, backgrounding, and logging.

bio-reproducer