The Agent Skills Directory

[COMMAND_EXECUTION]: Multiple scripts use subprocess.run to execute external commands and management tools.
scripts/optimize_skill.py and scripts/blind_comparator.py execute shell commands provided via arguments (e.g., --optimizer-cmd, --target-cmd) to interact with LLM providers.
scripts/trigger_test.py calls the claude CLI to judge triggering logic.
scripts/audit_skill_duplicates.py executes claude plugin list --json to inventory active plugins.
scripts/diff_skill_versions.py invokes git to compare skill versions across commits.
[DYNAMIC_EXECUTION]: The optimization loop in scripts/optimize_skill.py supports a custom verifier that executes a local script.
The verify_script function uses subprocess.run to execute a file path provided by the user, passing JSON-encoded task data to its stdin.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data that could influence agent behavior during optimization.
scripts/optimize_skill.py ingests task definitions from tasks.jsonl. This data is used to construct prompts and define assertions for LLM-based grading, creating an ingestion surface for untrusted instructions if the task set is sourced externally.
[EXTERNAL_DOWNLOADS]: The skill integrates with an external ecosystem for finding and adding skills.
The documentation and instructions guide the user to use the Skills CLI (npx skills), which fetches skill configurations and content from the https://skills.sh service and GitHub repositories (e.g., vercel-labs/add-skill).

skills-management