The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructs the user to bypass standard safety confirmations by enabling 'autopilot' mode with full permissions using the --allow-all and --max-autopilot-continues flags.\n- [COMMAND_EXECUTION]: The skill executes shell commands, including the GitHub CLI (gh) for issue management and sourcing a local script (source ~/.config/marketinc/board.sh) whose contents are outside the skill's direct control.\n- [DATA_EXFILTRATION]: The skill sends data to external GitHub repositories by creating and updating issues based on the project's progress and the user's initial objective.\n- [INDIRECT_PROMPT_INJECTION]: The orchestrator ingests untrusted user input (feature objectives) and propagates it through a chain of sub-agents (researcher, architect, implementer) without explicit sanitization or boundary markers.\n
Ingestion points: User-defined objectives and vision provided at the start of the pipeline.\n
Boundary markers: Absent; no instructions are provided to sub-agents to ignore potentially malicious instructions embedded in the objective text.\n
Capability inventory: The pipeline has extensive file-system access (read/write in docs/ and src/), network access (via the gh tool), and script execution capabilities.\n
Sanitization: No sanitization or validation of the input objective is performed before it is processed by the specialists.

orchestrator