The Agent Skills Directory

[PROMPT_INJECTION]: The generated prompt template includes instructions such as 'Do NOT ask for confirmation
proceed immediately with implementation'. This encourages the downstream agent to bypass standard human-in-the-loop safeguards during automated task execution.
[INDIRECT_PROMPT_INJECTION]: The skill represents an indirect injection surface because it constructs prompts by verbatim copying of content from potentially untrusted task files.
Ingestion points: Reads data from manifest.json, context.md, contracts/, and markdown files within the tasks/ directory.
Boundary markers: Present. The template uses explicit headers (e.g., === CONTEXT ===, === OBJECTIVE ===, === IMPLEMENTATION REQUIREMENTS ===) to delimit different data sections.
Capability inventory: The generated prompts authorize the downstream agent to use file editing tools, execute test suites (pytest), run linters (ruff, mypy, eslint), and perform git commits.
Sanitization: Absent. The skill instructions explicitly state to 'Copy content EXACTLY from task files' and 'do not summarize or rewrite', which allows instructions embedded in task files to be treated as high-priority prompt content.
[COMMAND_EXECUTION]: The skill documentation includes bash snippets for validating the integrity of generated files using grep and provides examples for manual agent execution using the claude CLI.

parallel-prompt-generator