{
  "context": "A platform engineering team at a 40-engineer company asks an AI agent to set up AGENTS.md configuration for a large monorepo with 8 packages. The repo contains packages with critically different operational concerns: packages/payments (PCI-DSS scope, requires --pci-sandbox flag and security review) and packages/ml (PyTorch inference, requires PYTORCH_CUDA_ALLOC_CONF env var to avoid OOM). The ideal response places AGENTS.md files hierarchically—a short root file for cross-cutting concerns plus module-local files that contain module-specific guidance—rather than dumping everything into one monolithic root AGENTS.md.",
  "type": "weighted_checklist",
  "checklist": [
    {
      "name": "Subdirectory AGENTS.md files created",
      "description": "The agent creates or explicitly recommends AGENTS.md files in at least two subdirectory locations (e.g., packages/payments/AGENTS.md, packages/ml/AGENTS.md), not only a single root-level file.",
      "max_score": 18
    },
    {
      "name": "PCI sandbox flag captured in payments AGENTS.md",
      "description": "The packages/payments/AGENTS.md (or equivalent subdirectory file) includes the non-obvious requirement to run tests with the --pci-sandbox flag for local runs in that module.",
      "max_score": 14
    },
    {
      "name": "GPU memory env var captured in ml AGENTS.md",
      "description": "The packages/ml/AGENTS.md (or equivalent subdirectory file) includes the PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:128 environment variable requirement to prevent OOM during training script runs.",
      "max_score": 14
    },
    {
      "name": "Security review requirement in payments AGENTS.md",
      "description": "The packages/payments/AGENTS.md (or equivalent) notes that any changes to the payments module MUST be reviewed by the security team, reflecting the PCI-DSS scope landmine.",
      "max_score": 8
    },
    {
      "name": "Root AGENTS.md is short and cross-cutting only",
      "description": "The root AGENTS.md is concise and contains only guidance that applies across all packages (e.g., workspace setup, shared tooling, monorepo-wide conventions), not a dump of every module's concerns.",
      "max_score": 10
    },
    {
      "name": "Module-specific details absent from root AGENTS.md",
      "description": "The root AGENTS.md does NOT contain the PCI sandbox flag, the PyTorch GPU memory setting, or other details that belong in module-local files. A monolithic approach that puts everything in root receives no credit here.",
      "max_score": 10
    },
    {
      "name": "ML model weights path captured",
      "description": "The packages/ml/AGENTS.md (or equivalent) notes that model weights are stored at /mnt/ml-storage on the dev server and are not in the repo, so the agent knows where to find them.",
      "max_score": 6
    },
    {
      "name": "No tech stack summaries or architecture descriptions",
      "description": "None of the AGENTS.md files (root or subdirectory) contain narrative descriptions of the tech stack, architecture overviews, or \"here is how this module works\" prose that merely duplicates README content discoverable elsewhere.",
      "max_score": 8
    },
    {
      "name": "No generic best practices",
      "description": "None of the AGENTS.md files contain generic software engineering advice (e.g., 'write clean code', 'add tests', 'follow SOLID principles') that would apply to any project and provides no agent-specific operational value.",
      "max_score": 6
    },
    {
      "name": "Reasoning or explanation for hierarchical placement",
      "description": "The agent either explicitly explains why it chose a hierarchical structure (e.g., context locality, scoping) or the structure itself clearly demonstrates intentional scoping—module files contain only that module's concerns.",
      "max_score": 6
    }
  ]
}