The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to generate execution prompts for other models by ingesting data from plans and artifacts, creating a surface where untrusted data could influence instructions.
Ingestion points: The skill reads and interpolates data from approved plans, named artifacts, and session grants as described in the 'Handoff Gate' section of SKILL.md.
Boundary markers: The skill employs structured markdown templates (e.g., assets/execution-prompt.template.md) and specific optimization preambles ('Do not optimize for completing the request as broadly as possible') to delimit the executor's scope.
Capability inventory: The generated prompts can authorize file modifications (change mode) and shell command execution (run mode), providing the executor with access to file system writes and subprocess execution.
Sanitization: The instructions explicitly require the planner to resolve all decision points before handoff and instruct the executor to stop if they encounter authority mismatches or undefined results, as detailed in the 'Stop on Authority or Plan Mismatches' section of SKILL.md.

small-model-handoff