The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains 'Mandatory Invocation' rules that instruct the agent to disregard its own assessment of task simplicity and force the execution of a multi-model deliberation workflow. Phrases such as 'Do not answer directly from your own inspection' and 'If the skill is triggered, the minimum required behavior is...' are used to override the agent's default decision-making logic.- [PROMPT_INJECTION]: The workflow ingests user-supplied tasks and interpolates them directly into prompts for four separate models (Opus, Sonnet, Gemini, and Codex) without utilizing boundary markers or sanitization. This creates an attack surface for indirect prompt injection where a malicious user prompt could influence the behavior of the planning seats.
Ingestion points: User-provided task prompt processed in Step 1 and passed to Step 2 (SKILL.md).
Boundary markers: Absent; the clarified prompt is described as the 'sole input' to the council seats.
Capability inventory: Uses spawn_agent, claude-runner, gemini-runner, and codex-runner, and executes finalized plans in Step 6 (SKILL.md).
Sanitization: Absent; no evidence of input escaping or validation before passing content to council seats.- [COMMAND_EXECUTION]: The skill supports an --auto mode which allows for the automatic execution of implementation plans generated by the Codex model. Because these plans are derived from untrusted user input through an LLM synthesis process, this represents dynamic execution of potentially untrusted commands.

council