The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill features a surface for indirect prompt injection as it ingests untrusted data from project documentation and configuration files during its 'build' and 'sync' operations. This risk is effectively mitigated by specific instructions that require the agent to 'distill' and 'summarize' information through its reasoning layer rather than performing direct, unfiltered transcription of the source material.\n
Ingestion points: SKILL.md (The 'build' command scans files like README.md, package.json, and pyproject.toml).\n
Boundary markers: The skill utilizes dedicated entry-point files (AGENTS.md, CLAUDE.md) and standard templates to separate project memory from general agent instructions.\n
Capability inventory: The skill is designed to create directories and write/rename markdown and YAML files within the project's repository.\n
Sanitization: Explicit guidelines in SKILL.md and references/templates.md enforce distillation, summarization, and source referencing to prevent the verbatim propagation of potentially malicious instructions.\n- [COMMAND_EXECUTION]: The skill manages files and directories to maintain persistent memory. It includes safety mechanisms, such as 'Overwrite' and 'Migration' guards, which require explicit user confirmation before any destructive actions—like renaming legacy configuration files or overwriting existing documentation—are performed.\n- [CREDENTIALS_UNSAFE]: The skill incorporates a robust 'No secrets' policy as a core design principle. It provides clear instructions to avoid storing credentials, private keys, tokens, or personally identifiable information (PII) within the memory structure, reducing the risk of accidental data exposure.

agent-memory