The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill encourages the installation and use of third-party plugins from unknown GitHub accounts (Lum1104/Understand-Anything and mksglu/context-mode). It provides specific commands to add these to the plugin marketplace and install them, which involves fetching and executing external code at runtime to perform semantic analysis.
[INDIRECT_PROMPT_INJECTION]: The skill reads the entire codebase as untrusted input to generate AGENTS.md, a file that dictates rules and guardrails for all other agents. This allows malicious code or comments in the scanned project to potentially influence or override the behavior of the agentic pipeline.
Ingestion points: Codebase files are ingested in bulk via understand-anything or native shell tools in SKILL.md (Steps 3A and 3B).
Boundary markers: The skill lacks delimiters or explicit instructions to ignore embedded prompts within the source code being analyzed.
Capability inventory: The skill possesses the capability to write to the repository root (AGENTS.md) and project configuration folders (.claude/pipeline/), and can execute local shell commands.
Sanitization: No sanitization, validation, or escaping is performed on the data extracted from the codebase before it is integrated into the instructions in AGENTS.md.
[COMMAND_EXECUTION]: The skill executes shell commands such as find, grep, cat, and git to analyze the project structure and content when specialized plugins are not available. While part of its core functionality, this provides a mechanism to read any file in the project.
[EXTERNAL_DOWNLOADS]: The skill references external, unverifiable plugins for code analysis and context management that are hosted on GitHub by non-trusted individual accounts.

scan