The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local Python scripts (validate_brief.py and validate_briefset.py) included in the skill package to perform structural validation of generated Markdown documents. It also leverages standard development tools like git and ripgrep for codebase exploration. These are legitimate operations performed within the local environment for the skill's primary purpose.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input (Stage 1) to generate work plans, which are then processed by sub-agents in verification stages (Stage 5.5, 5.7). This creates a potential surface where malicious instructions in the source input could attempt to influence the sub-agent. The skill implements the following mandatory evidence chain to mitigate this risk: (1) Ingestion points include user-provided PRDs and task notes. (2) Boundary markers are enforced via a strictly defined Markdown H2/H3 template. (3) The capability inventory includes local Python validation and sub-agent reconstruction passes. (4) Sanitization is performed through a 'blind' execution-reconstruction check and a content-level self-check (Stage 5.6) designed to ensure the generated brief is a faithful work instruction.

task-brief-creator-caveman