The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to execute shell commands to locate and search memory files. It uses a specific command pattern to resolve the memory directory based on the current working directory: MEMORY_DIR="$HOME/.claude/projects/$(pwd | sed 's|/|%2F|g; s|%2F|/|; s|^/||')/memory". It also uses grep to search within these files.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves processing data from MEMORY.md (which accumulates content from previous sessions, potentially including untrusted third-party data or malicious instructions) and graduating it into permanent instruction files.
Ingestion points: Data is read from the project's memory file located at $MEMORY_DIR/MEMORY.md (SKILL.md, Step 2).
Boundary markers: The skill does not define explicit delimiters or security warnings when writing the distilled rules to the target files.
Capability inventory: The agent can read files, execute local shell commands (grep, sed), and write to project-level rules (CLAUDE.md, .claude/rules/*.md) as well as global user configuration (~/.claude/CLAUDE.md).
Sanitization: The skill relies on the agent's internal "distillation" process (Step 4) to convert descriptive notes into prescriptive rules, but lacks formal validation to ensure malicious directives aren't carried over.
[PERSISTENCE]: By design, the skill allows the agent to modify its own future behavior across sessions by writing to instruction files. Modifying ~/.claude/CLAUDE.md effectively alters the agent's global personality or constraints for all projects belonging to the user.

promote