The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill builds a mental model by reading existing codebase subsystems during its 'Grounding' phase. This creates a surface where instructions embedded in processed source code or documentation could theoretically attempt to influence the agent's behavior. However, the skill employs a structured evaluation process, including rationale templates and design red flags, which serves as a natural constraint against arbitrary instruction following from data.
Ingestion points: Phase A: Grounding reads subsystems via the how skill; Phase B: Sketch consumes grounding artifacts.
Boundary markers: The skill relies on a structured rationale template (RATIONALE-TEMPLATE.md) and explicit design criteria (DESIGN-RED-FLAGS.md) to synthesize and validate architecture.
Capability inventory: The skill can invoke other tools (how, why, arena, principles, interrogate), write architecture rationales and sketches, and implement application code.
Sanitization: Content ingested from subsystems is not explicitly sanitized, but it is processed within a specific architectural context.
[DATA_EXPOSURE]: The skill tracks its progress and configuration using local files in ~/.local/state/ox/ and ~/.config/ox/. These locations are standard for application state and configuration management. The analysis found no evidence of sensitive data (such as credentials or keys) being stored in these files or transmitted to external domains.

architect