The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains explicit instructions to bypass user oversight and safety loops. In the 'Plan Goal Steer' section, it commands the agent to: 'Do NOT wait for user confirmation. Execute the full goal autonomously.' and 'Immediately refine the goal objective and enter PABCD orchestration.'
[COMMAND_EXECUTION]: The skill framework relies on the autonomous execution of shell commands (cli-jaw orchestrate A/B/C/D) to transition through work phases without human-in-the-loop approval. The text explicitly states that goal mode 'supersedes PABCD phase gates for autonomous execution,' meaning no user wait occurs between planning and execution.
[PROMPT_INJECTION]: The skill is highly vulnerable to Indirect Prompt Injection (Category 8). It is designed to ingest untrusted data from the codebase and then execute actions autonomously based on that data.
Ingestion points: During the 'Context Scan' phase, the agent reads project documentation (README, CLAUDE.md), git history, and source code.
Boundary markers: The instructions lack any requirement for boundary markers or sanitization of ingested content before processing it for planning.
Capability inventory: The skill has the capability to execute shell commands (cli-jaw), write to the filesystem (creating devlogs and artifacts), and perform git commits.
Sanitization: There is no evidence of sanitization or filtering of external content before it is interpolated into the autonomous planning and execution logic.

goal