The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides instructions for managing and optimizing context that may contain untrusted data. It explicitly addresses the risk of goal hijacking (ASI01) and context poisoning (ASI06) by recommending the use of boundary delimiters and ensuring that security-related 'trust wrappers' are excluded from semantic pruning processes.
[COMMAND_EXECUTION]: The skill utilizes file system tools (write_file, edit_file, create_file) to persist optimized context states, such as compact-working-state.json. These operations are transparent, well-documented in the 'Output Contracts' section, and serve the legitimate purpose of maintaining durable state across agent sessions.
[SAFE]: The instructions align with modern LLM provider features (e.g., Anthropic Prompt Caching, OpenAI Prompt Caching) and follow industry-standard security frameworks for AI, including a dedicated section for OWASP Top 10 for LLM Applications (ASI) guardrails.

context-optimization