The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses a broad attack surface for indirect prompt injection by design, as it ingests untrusted content from chat transcripts to generate persistent instructions.
Ingestion points: The skill reads JSONL transcripts from ~/.claude/projects/ to identify user patterns (File: SKILL.md, Step 1).
Boundary markers: The skill does not define specific prompt delimiters for the mined data, though it instructs the agent to "Cross-check across slices before elevating a signal" to filter noise.
Capability inventory: The skill can write new files to the filesystem (.claude/skills/), execute git commands, and create Pull Requests.
Sanitization: No explicit content sanitization or escaping is mentioned for the mined data before it is incorporated into the drafted skill.
[COMMAND_EXECUTION]: The skill utilizes local shell commands to manage version history and deployment workflows.
Evidence: Uses git log to determine the scope of history mining based on the last edit time (File: SKILL.md, Step 0).
Evidence: Instructs the agent to use git worktree and git commit to finalize and propose the generated skill (File: SKILL.md, Step 6).

automate-me