The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves processing external materials, such as research documents, user feedback, and real-world task samples, to generate new agent instructions.
Ingestion points: references/development.md and SKILL.md describe activities like "观察真实对象" (observing real objects), "调研作品、理论与实践" (researching works, theory, and practice), and collecting "用户反馈" (user feedback).
Boundary markers: The instructions explicitly require the use of "隔离的完整试验" (isolated full experiments) and "真实对象的隔离副本" (isolated copies of real objects) to ensure that untrusted data or experimental instructions do not impact production environments.
Capability inventory: The skill has the capability to generate and execute "候选 skill" (candidate skills) and "任务块" (task blocks) through model calls in L1 and L2 testing phases.
Sanitization: The methodology emphasizes process-based isolation and human review of the generated output rather than automated input sanitization.
[DYNAMIC_EXECUTION]: The skill is designed to dynamically generate and refine executable instruction blocks (skills) based on observed needs and theoretical input.
Evidence: SKILL.md defines a specific workflow for "组织执行" (organizing execution) where core logic is translated into "可独立执行的候选 skill" (independently executable candidate skills) that are iteratively tested and updated based on performance data.
[SAFE]: The skill instructions demonstrate defensive design by explicitly prohibiting the creation of "永久总控" (permanent controllers) or "默认定时运行" (default scheduled runs). It further states that security boundaries like permissions and data consistency are the responsibility of the "宿主与工具" (host and tools), and the skill must operate within those provided limits without attempting to bypass them.

skill-incubator