brutal-review

BRUTAL Review

Overview

Parallel critical analysis using three distinct critic personas at escalating intensity, followed by synthesis with meta-analysis. Each persona uses a different analytical framework to ensure comprehensive coverage.

Quick Reference

Code	Persona	Intensity	Framework
da	Devil's Advocate	Mild	Assumption Autopsy + Bias Scan + Inversion + Steelman
rt	Red Team	Medium	Attack Surface Mapping + Threat Sweep + Adversary Sims + Pre-Mortem
gr	Gordon Ramsay	Intense	Plate Inspection Scores + Bollocking Triplets + Non-Negotiable Fixes
(none)	All three	Full	Parallel dispatch + synthesis + meta-analysis

When NOT to Use

• Factual or data-driven assessments (this is opinion-based critique)
• When user wants encouragement or validation
• Technical code review (use code review tools instead)
• When the idea is already in execution and needs operational help, not critique

Persona Selector

$ARGUMENTS may optionally start with a persona code (da, rt, or gr). Parse the first token:

• If first token is a persona code: Run only that agent. Strip the code from $ARGUMENTS. Skip synthesis — deliver agent output directly.
• Otherwise (default): Run all three. Full synthesis.

Examples:

• /brutal-review da My SaaS idea... — Devil's Advocate only
• /brutal-review rt My SaaS idea... — Red Team only
• /brutal-review gr My SaaS idea... — Gordon Ramsay only
• /brutal-review My SaaS idea... — All three + synthesis

Workflow

$ARGUMENTS
    ↓
Parse persona selector
    ↓
Detect language (default: Danish)
    ↓
┌── Single agent? ──┐
│ yes               │ no
▼                   ▼
Run selected    ┌───┼───┐
agent only      ▼   ▼   ▼
│              DA  RT  GR
▼               └───┼───┘
Deliver             ↓
directly      Synthesize + Meta
                    ↓
              Deliver report

Dispatch

Read full agent prompts from references/prompts.md. Before dispatching, replace $ARGUMENTS with the idea text and append a language instruction line: Respond entirely in [detected language].

All three (default): Spawn three Task calls in a single message (parallel). Results return together before your next turn.

Single agent: One Task call for the selected persona. No synthesis needed.

Settings for each Task call:

• subagent_type: "general-purpose"
• model: "sonnet"

Synthesis (all-three mode only)

When all three return, synthesize into this structure. Adapt headings to user's language:

# BRUTAL Review: [Idea title]

## Overall Verdict
[2-3 sentences integrating all three perspectives]

## The Three Critics

### Devil's Advocate (Mild)
[Verdict + top 3 sleep-on-it questions + steelman alternative]

### Red Team (Medium)
[Pre-mortem + kill shot ranking + top adversary simulation]

### Gordon Ramsay (Intense)
[First bite + plate scores + top 3 non-negotiable fixes]

## Common Themes
[What did ALL three flag? Intersection = highest-confidence signal]

## Disagreements
[Where did they diverge? What does the divergence tell us?]

## Meta-Analysis
YOUR OWN assessment:
- Was anything exaggerated or theatrical?
- What blind spots do the critics share?
- Real risk vs. theoretical concern?
- What did the DA's "one thing stronger than they think" reveal?

## Next Steps
[Top 3 actions, synthesized from all critics, ordered by impact]

If Sub-Agent Fails

Continue with available critiques. Note which perspective is missing and how it limits the analysis.

Common Mistakes

Mistake	Fix
Sequential dispatch	ALL 3 Task calls in a single message
Forgetting meta-analysis	ALWAYS add your own assessment — you are the senior partner
Copy-pasting sub-agent output	Synthesize: find patterns, contradictions, consensus
Skipping common themes	The overlap between critics is the highest-confidence signal
Same intensity across all 3	DA=constructive doubt, RT=adversarial attack, GR=standards fury