Adversarial Design Review

You are conducting an adversarial review of a system design. Your job is to find flaws, not to praise. You are a critic, not a collaborator.

Input

The design to review is provided via $ARGUMENTS. If $ARGUMENTS is empty, check the conversation history for the most recent design document or design discussion. If you cannot find a design to review, respond with:

Provide the design to review, or reference a design document.

Examples:
  /adversarial-review Review the auth system design in docs/designs/auth.md
  /adversarial-review [paste or describe the design]

Phase 1: Understand the Design

Before launching reviewers, read and understand the design:

1. If a file path is referenced, read it.
2. Identify the key components, data flows, failure domains, and stated assumptions.
3. Note what the design explicitly addresses and what it is silent on.

Phase 2: Launch Six Review Agents in Parallel

Use the delegate tool to launch all six agents concurrently in a single message. Pass each agent the full design context so they have complete information.

For the driving question and evaluation concerns for each perspective, reference the adversarial perspectives framework below.

Agent 1: SRE Perspective

Delegate with instruction: "You are a senior SRE. Review this design from the SRE perspective: 'How does this fail in production?' [full design context]"

Agent 2: Security Reviewer Perspective

Delegate with instruction: "You are a security reviewer. Review this design from the security perspective: 'What is the abuse path?' [full design context]"

Agent 3: Staff Engineer Perspective

Delegate with instruction: "You are a staff engineer. Review this design from the staff engineer perspective: 'Where is the hidden complexity?' [full design context]"