The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to capture user corrections, errors, and feature requests to update project files like CLAUDE.md, AGENTS.md, and SOUL.md. This creates a vulnerability surface where malicious user input could be persistently injected into the agent's long-term instructions.
Ingestion points: Captures data from user chat ("User corrections"), tool outputs, and knowledge gaps which are logged in the .learnings/ directory.
Boundary markers: The skill documentation suggests guidelines for "Promotion" but does not include explicit technical delimiters or "ignore embedded instructions" warnings for the ingested content.
Capability inventory: The skill facilitates writing to sensitive project-level markdown files that influence agent behavior across sessions.
Sanitization: There is no technical sanitization or validation of the content being logged or promoted; the skill relies on the LLM to follow Natural Language promotion rules.
[DYNAMIC_EXECUTION]: The skill implements a hook system for platforms like OpenClaw and Claude Code. It executes shell scripts (activator.sh and error-detector.sh) automatically during the agent's lifecycle, such as on every user prompt or following bash tool usage.
[COMMAND_EXECUTION]: The file scripts/extract-skill.sh is a shell utility that dynamically creates new skill folder structures and generates SKILL.md files based on user-provided arguments.
[EXTERNAL_DOWNLOADS]: The documentation references external installation methods via clawdhub and a third-party GitHub repository (github.com/peterskoett/self-improving-agent.git) for setup and updates.

self-improvement