The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses local shell commands such as grep, sed, wc, and find to analyze and manage instructions and memory files within the project root and the user's home directory (~/.claude/). These operations are restricted to the local environment and are necessary for the skill's primary functionality.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from auto-captured memory (MEMORY.md) and tool output via a PostToolUse bash hook (error-capture.sh). This data is processed to generate suggestions for project rules, which could potentially be influenced by content within the captured data.
Ingestion points: Analyzes MEMORY.md from the project's memory directory and reads from the CLAUDE_TOOL_OUTPUT environment variable during the error capture process.
Boundary markers: None explicitly defined in the automation logic to separate user-provided instructions from auto-captured memory patterns during processing.
Capability inventory: The skill has the ability to read and modify project configuration files (CLAUDE.md, .claude/rules/*.md) and execute shell scripts as part of the Claude Code hook system.
Sanitization: The skill identifies patterns but does not explicitly sanitize the content of memory entries before interpolating them into suggested instructions or rule files.

self-improving-agent