The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data in the form of FASTA, FAA, and GenBank files, which creates a surface for indirect prompt injection attacks.
Ingestion points: The skill reads content from user-provided files (e.g., database.fasta, input_gb) in scripts/curate_fasta.py and the provided Biopython snippets. The agent is specifically instructed to sample raw headers using grep and head in SKILL.md.
Boundary markers: Absent. There are no explicit instructions or delimiters used when the agent is prompted to inspect or summarize the content of the input files to prevent it from following instructions embedded within FASTA headers.
Capability inventory: The skill possesses file system capabilities, including directory creation (Path.mkdir) and file writes (Path.write_text, Path.replace) via scripts/curate_fasta.py. It also enables the execution of local scripts using uv run.
Sanitization: The normalize_header function in scripts/curate_fasta.py provides some mitigation by replacing whitespace and special characters (excluding _, ., |, :, and -) with underscores, which sanitizes the data for the output files, but does not protect the agent during the initial inspection of raw input files.

bio-fasta-database-curator