The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reads user-provided or previously generated documentation files and passes extracted summaries to sub-agents. If these files contain malicious instructions, the sub-agents might be influenced to deviate from their documentation tasks.
Ingestion points: Reads from docs/01-需求调研/ (MVP方案.md, 需求调研报告.md, etc.) in Step 2.
Boundary markers: Absent. Summarized data is interpolated directly into the sub-agent prompt templates (e.g., {从步骤2提取的关键信息}).
Capability inventory: Access to Write tool for document creation and Agent tool for task delegation.
Sanitization: Absent. Extracted data is not sanitized or escaped before interpolation.
[COMMAND_EXECUTION]: The preamble executes a shell script via a relative path. This script resides outside the skill folder.
Evidence: bash "$(dirname "${BASH_SOURCE[0]}")/../../check-update.sh" in SKILL.md.

pm-docs