The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes memory entries from user-controlled files which could contain malicious instructions designed to influence agent behavior.
Ingestion points: Reads from MEMORY.md, CLAUDE.md, topic files, and rules files in project and user directories (e.g., ~/.claude/).
Boundary markers: The instructions do not specify markers to delimit processed data from instructions, increasing the risk that the agent might follow instructions embedded within a memory entry.
Capability inventory: The skill has the capability to read files in ~/.claude/**, edit project instructions in CLAUDE.md, and perform network requests using WebFetch.
Sanitization: No sanitization or validation of the content of the memories is performed before it is rewritten or promoted to other files.
[EXTERNAL_DOWNLOADS]: The skill fetches official documentation and guides to inform its refactoring logic.
Evidence: Fetches guides from code.claude.com and learn.chatgpt.com related to memory storage and path-specific rules.

refactor-memory