The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input in Step 2 via open-ended questions about skill performance. This data is processed in Step 3 to create synthesis summaries and improvement proposals. These proposals are then either filed as GitHub issues or passed to a skill-updater tool which modifies other skills. Without strict sanitization or boundary markers, a malicious user could provide feedback designed to trick the synthesizer into creating harmful instructions for other skills.
Ingestion points: User responses to "What worked well?" and "What was clunky or missing?" in Step 2.
Boundary markers: None specified for the synthesis or output stages.
Capability inventory: GitHub CLI (gh) for issue management; invocation of skill-updater for modifying skill files.
Sanitization: No explicit sanitization or validation of user feedback is mentioned before synthesis.
[COMMAND_EXECUTION]: The skill uses the GitHub CLI (gh issue list) in Step 5 to query for existing issues. It also implies the use of gh issue create (or similar) in Step 7 to file new feedback. This demonstrates the capability to interact with external code repositories and persistent issue trackers.
[DATA_EXPOSURE]: The skill reads sensitive local files, specifically ~/.claude/settings.json, to identify enabled plugins and locate skill files. Accessing configuration files in the user's home directory is a form of local data exposure, although it appears aligned with the skill's stated purpose of skill management.

skill-feedback-collector