The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/build_cohort_profile.py executes the pdftotext system binary via subprocess.run to extract text from PDF files. While it uses a list for arguments to mitigate shell injection, it executes a system utility on user-provided file paths.
[COMMAND_EXECUTION]: The skill instructions in SKILL.md direct the agent to execute shell commands for literature searches, such as bash "$EUTILS" search ... and running the build_cohort_profile.py script via python3.
[EXTERNAL_DOWNLOADS]: The script scripts/build_cohort_profile.py uses urllib.request.urlopen to fetch content from user-specified URLs. Although it checks for http:// and https:// protocols, this allows the skill to retrieve data from arbitrary external domains.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted external data from CSV codebooks, PDF documents, and remote URLs. This content is processed and presented to the agent as context (e.g., in cohort_profile.md and context_extract.md). If these sources contain malicious instructions, they could influence the agent's behavior.
Ingestion points: scripts/build_cohort_profile.py reads local files and fetches remote URLs provided by the user.
Boundary markers: The script wraps extracted context in markdown headers such as # Context: [source].
Capability inventory: The agent has access to Bash, Grep, Glob, Read, Write, and Edit tools, allowing for file system operations and shell command execution.
Sanitization: The script performs basic HTML-to-text conversion but does not specifically sanitize the resulting text for embedded instructions.

find-cohort-gap