The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes auto-generated memory files containing potentially untrusted data from tool outputs and user interactions, then promotes this data into permanent project instructions.\n
Ingestion points: The memory-analyst and error-capture tools ingest data from MEMORY.md, topic-specific markdown files, and direct Bash command output (error logs).\n
Boundary markers: The skill documentation does not mention the use of delimiters or 'ignore embedded instructions' prompts when analyzing and promoting memory entries.\n
Capability inventory: The skill is capable of modifying CLAUDE.md, creating new rule files in .claude/rules/, and generating new SKILL.md files, effectively altering the agent's long-term behavior.\n
Sanitization: No sanitization or validation mechanisms are described for the content being transitioned from scratchpad memory to enforced rules.\n- [REMOTE_CODE_EXECUTION]: The provided installation instructions involve executing a local shell script and adding a third-party marketplace source.\n
Evidence: Documentation mentions ./scripts/codex-install.sh --skill self-improving-agent and adding the alirezarezvani/claude-skills marketplace repository.\n- [COMMAND_EXECUTION]: The error-capture hook operates by monitoring the output of Bash tools to record insights, interacting directly with shell execution results.\n
Evidence: The hook is explicitly described as Monitoring command output for errors via PostToolUse for Bash.\n- [DYNAMIC_EXECUTION]: The skill-extractor agent generates new executable skill files (SKILL.md) based on patterns observed at runtime.\n
Evidence: The skill description for skill-extractor states it 'generates a complete skill: SKILL.md with proper frontmatter' from recurring solutions.

self-improving-agent