The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes historical agent transcripts to extract patterns, creating a surface for indirect prompt injection where malicious instructions in previous chats could influence the behavior of the subagents performing the mining.
Ingestion points: The skill reads JSONL transcript files from various local paths like ~/.claude/projects/, ~/.cursor/projects/, and ~/.pi/agent/sessions/.
Boundary markers: None. The instructions do not specify the use of delimiters or warnings to ignore embedded commands within the ingested transcript data.
Capability inventory: The skill utilizes shell commands through git log and invokes other tools like create-skill to create or modify local skill files.
Sanitization: None. The transcript content is analyzed without any validation or filtering to remove potential instructions.
[DATA_EXFILTRATION]: The skill is designed to read sensitive local files containing conversational history (agent transcripts) across multiple directories and platforms. While this is restricted to the user's workspace for the purpose of personalization, it involves accessing private interaction logs.

automate-me