The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's instructions in SKILL.md and its working examples utilize the Task tool with a specifically defined mode: "bypassPermissions" parameter. This configuration allows sub-agents (teammates) to execute their assigned roles autonomously without requiring user intervention or verification for individual actions, increasing the potential impact if a sub-agent's instructions are compromised.
[INDIRECT_PROMPT_INJECTION]: The skill architecture relies on a multi-step pipeline where data ingested by early-stage sub-agents is passed directly into the prompts of subsequent agents, creating an unvalidated trust chain.
Ingestion points: sub-agents such as the Analyst are tasked with reading and documenting local project files, RLS policies, and database schemas as seen in the Analyst (분석가) template in references/prompt-templates.md.
Boundary markers: The skill uses markdown headers like ## 맥락 and ## 목표 to structure the sub-agent prompts, but these are insufficient to prevent a sophisticated adversarial injection embedded within the project files being analyzed.
Capability inventory: Sub-agents are granted general-purpose capabilities and operate with bypassed permissions, enabling potential file writes or system modifications if the agent context is hijacked.
Sanitization: There is no evidence of sanitization, filtering, or instruction-ignoring wrappers applied to the variables (e.g., {result_1}, {architect_result}) when they are interpolated into follow-up agent prompts.

team-assemble