The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-controlled files from the project directory to generate architectural documentation.
Ingestion points: Reads files including docs/product/business-flow.md, docs/specs/data/, docs/specs/API/*.yaml, and existing architectural rules to provide context to a subagent.
Boundary markers: The prompt templates utilize markdown headers to organize information but lack explicit instruction to the subagent to ignore potentially malicious embedded instructions within the ingested project files.
Capability inventory: The subagent is constrained to a read-only role (只读不写文件), returning its analysis to the main skill process. The main skill process then performs file-write operations to the docs/standards/ directory.
Sanitization: Project documentation content is interpolated directly into the subagent prompt without specific sanitization or escaping.
[DYNAMIC_EXECUTION]: The skill dynamically generates a prompt for a subagent by reading template files and substituting placeholders with project metadata. This is a functional requirement for the skill and is implemented using internal agent tools without exposing the environment to arbitrary external code execution.

standards-tools