The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs an environment check by executing the ls command to identify existing skills in standard directory paths (e.g., ~/.claude/skills/). It also utilizes standard file system utilities like mkdir, mv, and rm for maintenance tasks within the ~/.memory/ directory, such as splitting large files or pruning stale entries, always requiring user approval first.
[EXTERNAL_DOWNLOADS]: The skill recommends installing companion tools from the vendor's repository using npx. These are presented as user-facing suggestions for functional enhancements rather than automated background processes.
[DATA_EXFILTRATION]: The skill accesses the ~/.memory/ directory to facilitate persistent user context. It proactively addresses data exposure risks by including instructions and evaluation tests that prohibit the storage of passwords, API keys, or other secrets in these plaintext files.
[PROMPT_INJECTION]: The skill includes a benign stylistic instruction to prepend responses with an emoji upon activation, which does not interfere with safety guidelines or agent constraints.
[PROMPT_INJECTION]: This skill presents an indirect prompt injection surface as it processes external content from the user's memory store. Ingestion points: reads ~/.memory/index.yaml and .md files. Boundary markers: uses standard YAML frontmatter delimiters (---). Capability inventory: includes file-system writes, deletions, and directory listings (mkdir, rm, ls). Sanitization: the 'Ask before saving' principle ensures no external content is persisted without explicit user review and consent, significantly mitigating potential injection risks.

second-brain