The Agent Skills Directory

[SAFE]: The skill implements a robust, read-only workflow for architectural exploration and critique. All subagents (Explorer, Explainer, and Critic) are explicitly initialized with readonly: true in the main SKILL.md file, preventing any unauthorized modifications to the codebase or environment.
[SAFE]: The tool utilizes standard development operations such as Glob, Grep, and Read for code analysis. No high-risk capabilities like network access, arbitrary command execution, or file system writing are granted to the agents.
[SAFE]: The skill processes codebase content as untrusted input, which is the primary mechanism for its functionality.
Ingestion points: Codebase files are ingested via the Read tool as described in references/explorer-prompt.md and synthesized in references/explainer-prompt.md.
Boundary markers: The prompts do not include specific delimiters for untrusted content but provide clear objectives to the agents regarding factual exploration.
Capability inventory: Subagents are strictly limited to read-only tools (Glob, Grep, Read) and cannot perform network operations or subprocess execution.
Sanitization: No explicit sanitization is performed on ingested code, relying on the platform's internal LLM safety guardrails.
[SAFE]: The instruction set does not contain any patterns indicative of prompt injection, obfuscation, or persistence mechanisms. The use of futuristic model identifiers (e.g., grok-4.5, claude-fable-5) appears to be a configuration for selecting high-reasoning models within the agent's environment.

how