The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reads from MEMORY.md, a file containing data gathered during AI sessions. This data is untrusted as it may include information from external code, issues, or pull request descriptions. The skill facilitates promoting this content into CLAUDE.md or .claude/rules/, which are permanent, enforced agent instructions.
Ingestion points: The skill reads content from the project memory directory: $HOME/.claude/projects/.../memory/MEMORY.md (Step 2).
Boundary markers: No specific boundary markers or "ignore instructions" warnings are used when reading or distilling memory content.
Capability inventory: The skill allows the agent to write new rules into CLAUDE.md and create/edit instruction files in .claude/rules/ (Step 5).
Sanitization: The agent is instructed to "distill" rules into an "imperative voice," but there is no mechanism to validate that the source memory hasn't been influenced by adversarial content encountered during a previous session.
[COMMAND_EXECUTION]: The skill executes local shell commands to navigate the file system and search for existing patterns in the agent's memory.
Evidence: Use of sed to construct the memory path from the current working directory and grep to search and display matching entries from the MEMORY.md file (Steps 2 and 6).

promote