The Agent Skills Directory

[PROMPT_INJECTION]: The skill performs Indirect Prompt Injection by ingesting untrusted data from past conversation transcripts.
Ingestion points: Reads agent-transcripts/ from the active workspace directory to identify user patterns.
Boundary markers: The skill does not define explicit delimiters for transcript data but implements a frequency filter, requiring patterns to appear in multiple slices to be considered high-confidence.
Capability inventory: Uses the create-skill tool to write files to the filesystem and utilizes git for branch and PR management.
Sanitization: Relies on human review at the end of the process and applies the 'unslop' prose discipline skill to the output.
[DATA_EXFILTRATION]: The skill includes explicit instructions to protect user privacy by limiting transcript mining. It forbids the agent from globbing across ~/.cursor/projects/*/, ensuring that private chats from unrelated workspaces are not accessed or processed.
[COMMAND_EXECUTION]: The skill uses local shell commands for standard repository management. This includes git log to determine the time of the last skill update and standard git workflow commands (worktree, commit, PR) to propose changes to the user for review.

automate-me