The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes job postings and search results from external sources, which constitutes a vulnerability surface for indirect prompt injection.
Ingestion points: Job posting data is ingested from external searches and stored in jobs.json for processing.
Boundary markers: Instructions in SKILL.md and references/profile-and-discovery.md explicitly mandate treating external content as untrusted data and warn against following embedded instructions or granting authority based on source content.
Capability inventory: The skill has capabilities to write files to a private workspace and execute a local companion script for state management.
Sanitization: Input data is strictly validated by the toolkit/scripts/job_state.py script, which enforces JSON schema checks, UTF-8 verification, and recursion depth limits.
[COMMAND_EXECUTION]: The workflow requires the execution of a local Python script (toolkit/scripts/job_state.py) included in the skill package.
The script is an original tool for offline state management that uses only standard libraries and avoids network connections or subprocess spawning.
It implements security hardening using POSIX O_NOFOLLOW and dir_fd primitives to prevent symlink attacks and path traversal during file operations.
File access is limited to explicitly approved private workspaces, and new files are created with restricted permissions (0o600).

omh-ai-job-search