The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local scripts and commands to perform metadata searches and verify implementation changes.
Evidence: In Phase 2, the skill is instructed to run python .codestable/tools/search-yaml.py --dir .codestable/compound --query "{能力关键词}".
Evidence: The skill uses git log to scan for code changes related to requirements.
[PROMPT_INJECTION]: An indirect prompt injection surface exists because the skill ingests untrusted external data that could influence the agent's behavior.
Ingestion points: User materials (oral feedback, product ideas, user feedback), project files (requirements/*.md), and existing feature descriptions in SKILL.md (Phase 2).
Boundary markers: No delimiters or explicit instructions to ignore embedded commands within ingested data are defined.
Capability inventory: The skill has file-writing capabilities (creating and updating .md files in requirements/) and command execution capabilities (running search scripts and git).
Sanitization: There is no evidence of input validation or sanitization for the materials processed during the drafting phase.

cs-req