The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to read external prompt files and user-provided requirements to refactor them into a DSL. This ingestion of untrusted data constitutes an indirect prompt injection surface. However, the skill incorporates strong mitigation strategies.
Ingestion points: Workflow @compose in SKILL.md reads user-specified targets and requirement files.
Boundary markers: The skill defines an @authority_boundary which explicitly mandates that the agent must treat targets as data/material and MUST NOT execute tasks described within the ingested material.
Capability inventory: The skill is capable of reading and writing/editing local files as part of its refactoring workflow.
Sanitization: The skill transforms the input into a formal DSL structure, which naturally limits the potential for raw instruction passthrough to the output if followed correctly.
[COMMAND_EXECUTION]: The skill has the capability to edit user-specified files. This is an intended feature for a refactoring tool and is constrained by rules requiring the preservation of the original intent and scope of the source material.

use-prompt-dsl