The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell and Python scripts included in its package, such as scripts/list-skills.sh, scripts/check_paths.py, and 07-脚本与工具/surface_candidates.py. These scripts are used for discovering other installed skills, validating configuration paths, and searching for content candidates.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted user content, including video topics, scripts, and business data, to generate prompts for other specialized AI agent skills. This creates an attack surface where malicious data could influence downstream agent behavior.
Ingestion points: Processes user-provided topics, draft scripts, and business scenarios in SKILL.md (Mode A/B) and the Onboarding mode.
Boundary markers: Employs recap sections and structured choice-based interaction menus to separate internal instructions from user-derived content.
Capability inventory: Executes subprocesses via bash and python3 to interact with the local filesystem and discover other skills.
Sanitization: Implements a comprehensive "Redline" system defined in references/家族公约.md and config.yaml to filter prohibited words and regulate claims (e.g., redlines.banned_words, redlines.sell_real).
[DYNAMIC_EXECUTION]: The skill uses scripts/list-skills.sh to dynamically scan directories such as ~/.claude/skills and ~/.agents/skills at runtime. It parses metadata from these external skill folders to populate a candidate list, which is then used to route user tasks.

zmm