The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to use shell commands such as find, grep, stat, and jq to locate and analyze transcript files across various project directories. These tools are used for discovery and parsing of local JSONL transcript files.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from past interaction transcripts (JSONL files). This creates an attack surface where instructions embedded in past conversations could potentially influence the subagents during the pattern extraction process. Evidence: Step 4 (Dispatch parallel scan) and Step 8 (Walk through findings). The ingestion points are past session files, and while the skill includes logic for schema verification and filtering, it lacks explicit prompt boundary markers for the untrusted content. However, the human-in-the-loop review (Step 8) provides a significant layer of sanitization.
[DYNAMIC_EXECUTION]: The skill is designed to generate and write new executable content, including slash commands (markdown files in commands/ directories) and new AI agent skills (SKILL.md files). Evidence: Step 9 (Handle each decision
Execute promotion). This allows for the creation of new functional capabilities based on learned data.
[PERSISTENCE]: The skill modifies persistent configuration files such as ~/.claude/CLAUDE.md and project-specific CLAUDE.md or MEMORY.md files, which affects the agent's behavior in future sessions. Evidence: Step 9 (Execute promotion). This is the primary purpose of the skill but involves modifying core agent behavior rules.

audit-transcripts-for-learnings