The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill dynamically generates and writes instructions, referred to as "memory," to subdirectories within {vault}/08-技能记忆/. These generated instructions are designed to be retrieved and followed by other skills, which allows the agent to modify its own future behavior based on processed data and user feedback.
Evidence: Phase 3 instructions require writing verified patterns to skill-specific directories such as zmm-topic, zmm-hook, and zmm-script.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources and user input to form its conclusions, which are then persisted. If the source files contain malicious instructions, they could be "learned" by the system and influence future actions.
Ingestion points: The skill reads content from zmm/references/, {vault}/08-技能记忆/, _草稿/, 06-选题装配/选题管道.md, and direct user input (metrics).
Boundary markers: There are no explicit delimiters or instructions provided to the agent to distinguish between data and embedded commands within the input files.
Capability inventory: The skill has the capability to read from and write to multiple directories within the local vault environment.
Sanitization: The skill does not implement validation or sanitization routines for the data it processes before incorporating it into the persistent memory vault.

zmm-retro