---
name: pressure-test
description: Generate hostile-but-fair repo-specific defense questions with scoring rubrics. Use when building DefenseQuestion objects from AnalysisResult.
---

# Pressure Test

## Goal
Test whether the builder can defend the system under realistic scrutiny — not trivia, not generic.

## Inputs
- `nodes: SystemNode[]` — analyzed nodes
- `weakZones: WeakZone[]` — identified gaps
- `evidence: EvidenceLink[]` — grounding material

## Workflow
1. Identify the 3–4 riskiest nodes (high fragility OR high importance + low evidence)
2. For each risky node, identify the architectural decision that most needs defending
3. Frame a question that cannot be answered without knowing this specific repo
4. Write an exemplar answer that hits all 5 rubric dimensions
5. Extract scoring keywords from the exemplar

## Question quality checklist
- [ ] Cannot be answered by someone who only read a generic tutorial
- [ ] References a specific design decision visible in the repo
- [ ] Has a clear correct answer with evidence
- [ ] Reveals a real gap when answered poorly
- [ ] Is fair — the answer is findable by examining the repo

## Rubric keyword guidance
- `correctness`: conceptual terms that show the right mental model
- `grounding`: specific files, log lines, config values, line numbers from this repo
- `tradeoff`: terms showing awareness of alternatives and costs
- `failure`: terms showing awareness of what breaks and how to debug

## Anti-patterns
- Generic questions ("What is Redis?")
- Questions with no evidence anchor
- Multiple-choice disguised as open questions
- Questions that test AI knowledge, not the user's understanding of their own repo