The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes metadata, such as descriptions and trigger phrases, from other installed skills found in the .claude/skills/ directory. This creates a surface where a malicious or compromised skill could provide instructions designed to influence the agent when it later reads the generated guidance in CLAUDE.md.
Ingestion points: The scripts/scan-skills.ts script reads SKILL.md files from the .claude/skills/ directory.
Boundary markers: The generated output is wrapped in  comments, but individual skill descriptions are interpolated into the guidance section without specific delimiters or warnings to ignore embedded instructions.
Capability inventory: The skill uses Deno scripts with read access (--allow-read) and instructs the agent to write the generated content to CLAUDE.md or AGENTS.md.
Sanitization: The scripts/generate-guidance.ts script performs text truncation and first-sentence extraction to create concise summaries, but it does not include security-focused sanitization or validation of the ingested text content.

skill-integrator