The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from local transcript files (~/.claude/projects/ workspace-encoded path) to derive instructions for new agent skills.\n
Ingestion points: Conversation transcripts are read from ~/.claude/projects/<encoded-cwd>/*.jsonl to identify recurring user patterns and working conventions.\n
Boundary markers: The skill lacks explicit prompt delimiters for parsing transcripts, but it mitigates risks by requiring signals to appear in at least two different history slices to be considered high-confidence.\n
Capability inventory: Includes reading transcript files, writing new skill files to ~/.claude/skills/, and executing git commands (log, worktree, commit) for PR creation.\n
Sanitization: The skill employs the unslop skill for prose discipline and mandates a human-in-the-loop review process using the AskUserQuestion tool before finalizing the draft.\n- [COMMAND_EXECUTION]: The skill executes git commands (git log, git worktree, git commit) to check history and manage the lifecycle of the created skill via pull requests. These operations are restricted to the local repository context and follow standard developer workflows.

automate-me