The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill automatically updates the CLAUDE.md file based on user corrections and expressed preferences. This creates an attack surface where malicious instructions provided by a user or encountered in processed data can be persisted in the project's documentation, potentially influencing the agent's behavior and constraints in subsequent sessions.
Ingestion points: User preferences, corrections, and discovered patterns identified in SKILL.md.
Boundary markers: Absent. There are no delimiters or explicit instructions to treat user-provided preferences as untrusted content before writing them to the file.
Capability inventory: The skill performs file read and write operations to maintain the CLAUDE.md document.
Sanitization: Absent. The skill instructions encourage direct recording of user-provided statements as factual preferences.
[DATA_EXPOSURE]: The skill's instructions and the template in references/claude-md-template.md encourage recording environment variables, deployment patterns (including SSH details), and external integration details. If CLAUDE.md is committed to version control, which is the intended use for project conventions, it leads to the exposure of infrastructure details and potential credentials.
[PERSISTENCE]: The skill is explicitly designed to maintain instruction persistence across sessions by updating a project-level file. This mechanism allows for the creation of 'permanent' overrides to agent behavior that persist beyond the current conversation history.

project-conventions