The Agent Skills Directory

[PROMPT_INJECTION]: The skill creates a persistent feedback loop where the agent modifies its own instruction set based on observations and user feedback.
Ingestion points: Data is collected from user corrections, error messages, and tool outputs (accessed via the CLAUDE_TOOL_OUTPUT environment variable in scripts/error-detector.sh).
Boundary markers: Absent. The skill does not use specific delimiters or instructions to ignore potential malicious payloads within the captured text.
Capability inventory: The skill can modify core project instruction files (CLAUDE.md, SOUL.md, AGENTS.md) and generate new skill structures using the scripts/extract-skill.sh utility.
Sanitization: Absent. There are no mechanisms in the logging or promotion logic to filter, escape, or validate the content against instruction injection attempts.
[COMMAND_EXECUTION]: Includes several shell scripts (activator.sh, error-detector.sh, extract-skill.sh) intended to be executed by the agent's host environment.
extract-skill.sh automates the creation of directory structures and files. It includes basic regex validation for the skill name to mitigate path traversal risks.
[EXTERNAL_DOWNLOADS]: Documentation provides installation instructions that reference a third-party registry (clawdhub.ai) and a personal GitHub repository (github.com/peterskoett/self-improving-agent).

self-improvement