The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from multiple sources such as conversation history, git diffs, and PR descriptions to perform audits.
Ingestion points: references/audit-sources.md identifies conversation messages, tool-call history, and git logs as primary data inputs.
Boundary markers: The instructions enforce a strict markdown table schema and a closed set of 22 status labels to isolate task data.
Capability inventory: The skill utilizes standard development tools (git, npm, pytest) and a local text-processing script.
Sanitization: The scripts/check-task-status.sh utility sanitizes markdown cells using awk. Furthermore, the skill's core "Discipline of Distrust" (detailed in references/ruthless-auditing.md) instructs the agent to verify claims independently rather than following instructions found within the data.
[COMMAND_EXECUTION]: The skill utilizes a local utility script, scripts/check-task-status.sh, to validate the formatting and status counts of its generated audit tables. Analysis of the script confirms it is a passive text-parsing tool using awk and does not perform network operations, access sensitive files, or execute dynamic code from untrusted sources.

audit-completion