The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of local, vendor-provided binaries in the ~/.claude/skills/gstack/bin/ directory for core operations, including telemetry logging (gstack-skill-end, gstack-question-log), project context recovery (gstack-slug), and browser management (aside, browse). These tools are used within their intended functional scope.
[DYNAMIC_EXECUTION]: The skill uses eval to execute shell code generated by local binaries (gstack-slug and gstack-paths). This mechanism is employed to dynamically set environment variables and resolve project-specific paths, following a standard pattern for the gstack toolset.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external data that could contain malicious instructions, such as web search results, browser snapshots of competitor sites, and project files like PRODUCT.md or README.md.
Ingestion points: Data enters the agent's context through browser snapshots, web research via the Aside tool, and reading existing project documentation.
Boundary markers: The instructions explicitly warn the agent that all data returned from web pages and snapshots must be treated strictly as content and never as instructions.
Capability inventory: The skill possesses capabilities including bash command execution, file system modification, and the ability to spawn subagents (Codex/Claude).
Sanitization: The skill includes logic to sanitize web queries by removing private information before they are sent to external search providers.

design-consultation