The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes standard Unix utilities (bash, awk, jq, sed, tar, zip) for log analysis, file processing, and archive creation. These are standard operations for a diagnostic tool and are implemented with appropriate safety measures such as variable quoting and reading files in small chunks.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it ingests and processes untrusted session transcripts produced by other agents. This risk is effectively mitigated by references/context-safety.md, which enforces strict data intake limits (e.g., measuring file sizes, extracting only specific line ranges, and truncating fields) and prompts/scrub.md, which redacts sensitive information.
[DATA_EXFILTRATION]: The skill includes a workflow for exporting session data into bundles for bug reporting. This process is secured by multiple approval gates: the human partner must approve the redaction log, the file list, the exact text of any GitHub issue, and the final archive creation. A mandatory "scrub-audit" step performed by a subagent ensures the redaction policy is correctly applied before any data leaves the local environment.
[EXTERNAL_DOWNLOADS]: The skill interacts with the official GitHub API (api.github.com) and the GitHub CLI (gh) to search for and create issues within the author's repository (obra/superpowers). These operations are transparent, restricted to the official vendor domain, and require explicit user approval for any outgoing content.

diagnosing-superpowers