The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes git rev-parse --show-toplevel to dynamically resolve the AARON_SKILLS_ROOT path. This is a standard development practice for locating configuration files and internal scripts within a git repository.
[EXTERNAL_DOWNLOADS]: The skill references the author's homepage and repository at github.com/aaron-he-zhu/aaron-marketing-skills. As these are resources from the verified skill author, they are considered safe vendor resources.
[DATA_EXFILTRATION]: The skill reads configuration and benchmark files from parent directories (e.g., ../../../references/auditor-runbook.md). This access is consistent with a modular repository structure where shared references are stored centrally.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes user-provided narrative and experiment data. However, it implements a structured auditing framework (TALE) and explicit instructions to prevent the agent from mutating source data or executing autonomous publishing actions, which serves as an effective boundary:
Ingestion points: Reads user-supplied 'canon', 'surfaces', or 'message experiments' for auditing.
Boundary markers: Explicitly instructs the agent that its role is to 'judge' and that 'authoring/fixing belongs to Trace/Architect/Land skills'.
Capability inventory: Limited to reading references and writing audit artifacts to a specific local memory path (memory/audits/narrative/).
Sanitization: Uses typed observation vocabularies and fixed scoring constructs to ensure results are based on objective criteria rather than embedded instructions.

narrative-quality-auditor