The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates a workflow where external task descriptions are interpolated into subagent prompts, creating a surface for instructions embedded in data to influence agent actions.
Ingestion points: The implementer-prompt.md template takes the full text of a task from a plan file (e.g., created by superpowers:writing-plans) and pastes it into the implementer's prompt. The spec-reviewer-prompt.md template ingests both the original requirements and the implementer's output report.
Boundary markers: While the templates use Markdown headers like ## Task Description and ## What Was Requested to delineate sections, they do not utilize unique delimiters or explicit instructions to the subagent to ignore any contradictory instructions contained within the ingested task text.
Capability inventory: The 'implementer' subagent is a general-purpose agent tasked with writing code, creating and running tests, and performing git commits. This provides a significant capability surface if a malicious task description were to be processed.
Sanitization: No sanitization, escaping, or validation is performed on the task text before it is injected into the subagent prompts.

subagent-driven-development