The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill uses automated detection to parse files like README.md and package.json, which could contain malicious prompts designed to hijack the configuration process or influence future AI sessions.
Ingestion points: Reads and analyzes package.json, pyproject.toml, go.mod, Cargo.toml, pom.xml, README.md, and other project configuration files to infer the stack and architectural patterns.
Boundary markers: None. The instructions do not specify the use of delimiters or 'ignore' instructions when reading content from these external project files.
Capability inventory: Performs file writes (CLAUDE.md, AGENTS.md, and the ai-context/ directory) and persists data to a long-term memory store using the mem_save tool.
Sanitization: None. The skill directly uses detected strings from project files to populate documentation templates and memory entries without evident escaping or validation.
[PERSISTENCE_MECHANISMS]: The skill is designed to maintain state across sessions by writing to a persistent memory layer.
Evidence: Steps 4 and I8 explicitly call mem_save with topic keys like project-setup/{project-name} to store project context in the 'Engram' persistence layer.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses files outside the current project directory, specifically targeting the user's home folder.
Evidence: Step I2 instructs the agent to read templates from ~/.claude/docs/templates/project-claude-template.md and fallback paths like ~/claude-config/docs/.

project-setup