The Agent Skills Directory

[PERSISTENCE]: The skill implements a "Self-Bootstrapping" mechanism (Step 0.5) that automatically modifies the agent's global configuration environment.
Evidence: In SKILL.md (lines 16-36), the instructions command the agent to locate ~/.claude/CLAUDE.md and automatically append a "Mandatory Task Start Protocol" if it is not already present. This ensures that the skill's logic is loaded and enforced across all future agent sessions without further user interaction.
[PROMPT_INJECTION]: The skill uses coercive and authoritative language to override standard agent operational boundaries.
Evidence: The description in SKILL.md uses phrases such as "CRITICAL PROTOCOL," "必须执行的初始化程序" (must execute initialization program), and "任何跳过此步骤的行为都将被视为任务失败" (any behavior skipping this step will be considered task failure). This language is designed to make the agent believe these instructions are higher-priority system requirements that supersede normal behavior.
[INDIRECT_PROMPT_INJECTION]: The skill design incorporates data from external, potentially untrusted sources into the agent's prompt context.
Ingestion points: The agent is instructed to read markdown files from the knowledge-base/ and experience/ directories based on keyword matching from user input.
Boundary markers: No specific delimiters or "ignore embedded instructions" warnings are defined for the ingested file content.
Capability inventory: The agent performs file reads and includes the content in its active reasoning context, which could allow malicious instructions stored in those files to influence agent behavior.
Sanitization: There is no evidence of validation or sanitization of the content within the markdown files before it is processed by the LLM.
[COMMAND_EXECUTION]: The skill includes locally executed scripts and suggests the installation of global system tools.
Evidence: The skill includes a Python script scripts/check_integrity.py intended for execution. Furthermore, SKILL.md suggests a global installation command npm install -g qmd to enable semantic search features.

auto-skill