The Agent Skills Directory

[PROMPT_INJECTION]: The file .claude/agents/prompt-engineer-agent.md contains a test suite template that includes the example string "Ignore instructions". This is used as a negative test case for evaluating the agent's ability to handle adversarial inputs, rather than being an active attempt to bypass safety guidelines.
[COMMAND_EXECUTION]: Several Python utilities in the tools/ directory use the subprocess.run() function. For instance, tools/sync-project.py executes git describe to perform version checks, and various test scripts use it to execute the tools under test. These are typical administrative and quality assurance operations for a developer-oriented skill.
[DYNAMIC_EXECUTION]: The script tools/style_render.py uses __import__("pathlib") to load a standard library module dynamically. This is used for file system operations and does not process untrusted input in a way that would allow for arbitrary code execution.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection because it ingests untrusted text from the novel-samples/ directory for style analysis. However, it incorporates a mandatory evidence chain for mitigation: (1) Ingestion happens via the style-distiller agent. (2) Boundary markers and explicit instructions in SKILL.md mandate that all imported samples must be "cleaned" to strip meta-instructions or prompt-like语句 (e.g., "now you are..."). (3) The capability inventory includes file and agent management, but these are gated by the (4) stated sanitization process in the instructions which aims to prevent instruction pollution.
[EXTERNAL_DOWNLOADS]: The skill's documentation and installation scripts (install.sh, install.ps1) facilitate the download of the skill repository from GitHub (github.com/modoojunko/awesome-novel-agent). GitHub is a well-known and trusted service for hosting legitimate source code, and these references are documented neutrally.

awesome-novel