The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources including GitHub PRs, Jira issues via sibling skills, local CSV files, and Tech Breakdown documents.
Ingestion points: External data is ingested via gh pr view, gh pr diff, gh api, and the bitwarden-atlassian-tools:researching-jira-issues skill.
Boundary markers: The instructions explicitly command the agent to treat content from Jira, Confluence, PRs, and CSVs as untrusted data.
Capability inventory: Uses Read, Write (to local cache), Grep, Glob, Bash (restricted to gh and git commands), and Skill (Jira).
Sanitization: Instructs the agent to ignore imperative text within read content and flag suspicious payloads as CWE-1427 concerns, mitigating the risk of obeying hidden instructions.
[COMMAND_EXECUTION]: The package includes a Python utility (evals/run_real_eval.py) for developers to benchmark the skill's natural language triggering.
The script uses subprocess.Popen to invoke the claude CLI for automated testing.
It includes safety features such as EXEC_TOOLS and READ_ONLY_BASH whitelists to terminate agent processes if they attempt to perform heavy or unauthorized work during the evaluation of non-triggering queries.

assessing-test-coverage