lev-os Multi-Model Triangulation
Core Insight: Different models have different blind spots. Consensus = confidence.
How It Works
You can't directly call other models. Instead:
- I generate copy-paste prompts for you
- You paste into Codex/Gemini/Grok/etc
- You return their responses to me
- I synthesize into unified recommendation
Claude → generates prompt → You → paste to Model B → You → paste response back → Claude synthesizes
Quick Start
Tell me what you want triangulated:
Triangulate: [topic]
Context: [relevant details]
Models to use: [Codex, Gemini, Grok, or "all available"]
I'll generate the prompt(s). You copy-paste and return results.
Ready-to-Copy Prompts
Idea Evaluation
# COPY TO [Model Name]:
Evaluate these ideas. Score 1-10 on Quality/Utility/Feasibility/Risk:
1. [IDEA 1]
2. [IDEA 2]
3. [IDEA 3]
For each: scores, one-sentence rationale, final ranking.
Be critical—don't just agree.
Code Review
# COPY TO [Model Name]:
Review for bugs/security/improvements:
```[lang]
[CODE]
Categorize: Critical (must fix), Important (should fix), Suggestions. Overall score: X/10. Be thorough.
### Architecture Decision
COPY TO [Model Name]:
Choosing between: A: [Option A] B: [Option B] C: [Option C]
Evaluate: complexity, maintainability, performance, scalability. Recommend ONE with reasoning. Be opinionated.
More prompts: [PROMPTS.md](references/PROMPTS.md)
---
## Synthesis Template
After I receive responses from multiple models:
```markdown
## Triangulation: [Topic]
### Consensus (High Confidence)
- [Points ALL models agree on]
### Divergence (Investigate)
| Topic | Claude | Model B | Model C |
|-------|--------|---------|---------|
| [X] | [view] | [view] | [view] |
### Unique Insights
- **Claude:** [unique point]
- **Model B:** [unique point]
### Recommendation
[Synthesized recommendation]
### Confidence: [High/Medium/Low]
Model Strengths
| Model | Strengths | Best For |
|---|---|---|
| Claude | Nuance, safety, writing | Complex reasoning, docs |
| GPT/Codex | Code generation, breadth | Implementation details |
| Gemini | Multimodal, current data | Visual, recent events |
| Grok | Unconventional takes | Creative alternatives |
Tip: For security reviews, use ALL models. For routine code review, 2 is enough.
When to Triangulate
| Decision Type | Triangulate? | Why |
|---|---|---|
| High-stakes architecture | Yes | Hard to reverse |
| Security review | Yes | Blind spots are dangerous |
| Code review (routine) | Maybe | 1-2 models sufficient |
| Quick question | No | Overhead not worth it |
| Creative brainstorming | Yes | Different perspectives |
Rule: If hard to reverse or high-impact, triangulate.
Anti-Patterns
| Don't | Do |
|---|---|
| Ask vague questions | Specific, structured prompts |
| Accept first answer | Get 2-3 perspectives |
| Ignore disagreements | Investigate WHY models differ |
| Weight all equally | Consider model strengths |
| Skip synthesis | Always produce unified view |
Integration
Script Helper
# Generate formatted prompt for ideas
./scripts/format-prompt.py idea "Idea 1" "Idea 2" "Idea 3"
# For code (reads from stdin)
cat code.py | ./scripts/format-prompt.py code
# For architecture
./scripts/format-prompt.py arch "Use Redis" "Use PostgreSQL"
With Other Skills
| Combine with... | For... |
|---|---|
| ux-audit | Get multiple UX perspectives |
| multi-pass-bug-hunting | Cross-validate bug findings |
| idea-wizard | Score generated ideas |
References
| Topic | File |
|---|---|
| Full prompt library | PROMPTS.md |
| Real examples | EXAMPLES.md |
Scripts
| Script | Purpose |
|---|---|
scripts/format-prompt.py |
Generate copy-paste ready prompts |