The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from tool outputs (specifically file edits) to identify patterns, which presents a vulnerability to indirect prompt injection.
Ingestion points: The hook-templates/observer-hook.cjs.template file captures the resultPreview from tools like Edit and Write and stores them in .claude/memory/observations/.
Boundary markers: Observations are stored in structured JSON, but the captured tool output content lacks specific delimiters or warnings to prevent the analyzer from obeying instructions embedded within that data.
Capability inventory: The skill possesses the capability to write new instruction files to .claude/rules/ and .claude/skills/, which influence the agent's future behavior.
Sanitization: The scripts/validate-content.js script only validates YAML frontmatter and minimum length; it does not perform content sanitization or security scanning of the generated rules.
[DYNAMIC_EXECUTION]: The skill is designed to dynamically generate and install new instructions (rules and skills) for the agent environment.
Generation Mechanism: The agent-prompts/skill-generator.md prompt facilitates the creation of markdown-based instructions from analyzed session data.
Execution Context: The resulting files are saved to the agent's configuration directories (.claude/rules/), where they are automatically interpreted and applied in future sessions, effectively modifying the agent's behavioral logic at runtime.

continuous-learning-v2