Parallaxed User Guide
Everything you need to know about adversarial AI consultation.
What is Parallaxed?
Parallaxed is an adversarial AI consultation system. Instead of asking one AI for an answer, Parallaxed sends your challenge to multiple AI agents from different providers (Claude, GPT-4, Gemini, Mistral, and more). Each agent analyzes your challenge independently, then they cross-examine each other's positions. Finally, a mediator synthesizes everything into a structured deliverable.
The key insight: disagreements between agents are the most valuable output. When agents agree, that's expected. When they disagree, that's where blind spots, risks, and non-obvious solutions live.
Your First Consultation
- Type your challenge in the prompt bar at the bottom of the screen. Be specific — "Should I use microservices or a monolith for my e-commerce platform?" is better than "Help me build something."
- Choose a Forge Mode (or let the system suggest one). Adversarial is the default and works for most cases.
- Select a Mission from the sidebar if you want pre-configured agent teams. "Build" is great for implementation plans, "Research" for deep analysis.
- Click CONVERGE (or press Ctrl+Enter) to start the forge.
- Watch the activity log as agents analyze, cross-examine, and the mediator synthesizes.
- Read your report — it expands to full screen when complete. Look at the Disagreements Resolved section first — that's the gold.
Interface Overview
The interface has three main areas:
- Left Sidebar — Configuration: mission, forge mode, agent assignments, Intellinomics mode, depth, Diverge toggle, web search toggle, mediator provider.
- Center Panel — The activity log during forging, and the full-screen report after convergence.
- Right Sidebar — Project settings, file attachments, prompt templates, past reports.
The 3-Phase Protocol
Every consultation follows three phases:
Phase 1: Independent Takes
Each agent receives your prompt and analyzes it independently. They don't see each other's responses. This prevents anchoring bias — each agent forms their own opinion from scratch.
Phase 2: Cross-Examination
Agents see digests of each other's key claims and must respond. They're required to: quote and challenge at least 2 specific claims from other agents, concede 1 point that changed their mind, and identify 1 unaddressed risk. This phase is where the real value emerges.
Note: In Blind Review and Socratic modes, this phase is skipped — agents never see each other.
Phase 3: Mediator Synthesis
A mediator (typically Claude Sonnet or the provider you choose) reads all agent outputs and produces a structured deliverable. The mediator must attribute every position to its source agent, resolve every disagreement with explicit reasoning, and produce actionable next steps.
Forge Modes
Five modes change how agents interact, not just what they say:
⚔ Adversarial
Default mode. Agents actively challenge each other. Best for decisions, plans, and designs where you need blind spots surfaced.
◈ Pre-Mortem
Half the agents build the failure case, half build the success case. Best for launches, bets, and major decisions.
? Socratic
Agents may ONLY ask questions — no answers. Best for early-stage exploration when you don't know what you don't know.
◎ Council
Each agent represents a different stakeholder (end user, investor, regulator, competitor). Best for product and ethics decisions.
◐ Blind Review
Agents never see each other's output — pure independent analysis. The mediator highlights where they converged without coordination (high-confidence signals).
Q-Score Quality Measurement
Every consultation gets a Q-Score — a quality grade across 5 dimensions:
| Dimension | What It Measures | Why It Matters |
|---|---|---|
| Divergence | How much agents disagreed | High divergence = multiple genuine perspectives explored |
| Depth | Thoroughness of analysis | Deep analysis catches issues shallow analysis misses |
| Coverage | Breadth of topics addressed | Nothing important was left out |
| Resolution | Quality of mediator synthesis | Disagreements were resolved with reasoning, not hand-waving |
| Actionability | Concrete next steps produced | Output you can actually execute on |
Grades range from A+ (exceptional) to F (failed). Target: B+ or higher. If you consistently get below B, try deeper mode, more agents, or enable Diverge.
Agents & Roles
Parallaxed has 33 agent roles across 6 categories. Each role has a specialized system prompt that shapes their analysis perspective:
- Core Engineering: Dev, Frontend, Backend, Mobile, Fullstack
- Architecture: Architect, API Designer, Systems, UX, Creative
- Quality: QA, Security, Performance, Accessibility, Code Reviewer
- Infrastructure: Ops, DevOps, Cloud, Data, DBA
- Strategy: Research, Product, Business Analyst, Tech Lead, Critic
- Specialized: AI/ML, Docs, Compliance, Cost Analyst, Patent Agent, Brain Cramp, Brainstorm
Any role can be powered by any of the 14 supported AI providers. The Provider Fitness Matrix automatically assigns the best provider to each role based on 10 cognitive dimensions.
Intellinomics
"Intelligent selections concerning the economics of Vibe Coding." Three modes that control which AI providers and models are used:
- Thrifty — Budget providers only (Groq, DeepSeek, Together). Fastest, cheapest. Good for quick explorations. No web search.
- Balanced — Mix of premium and budget providers. Best value. Standard web search.
- Unleashed — Top-tier models everywhere (Claude Sonnet, GPT-4o, Gemini Pro). Deep web search with refinement rounds. Maximum quality, highest cost.
Diverge Enhancer
When enabled, Diverge takes your raw prompt and enhances it before agents see it. It adds structure, specificity, and Q-Score-aware framing. The enhanced prompt is shown for your review — you can accept, edit, or skip it.
Impact: Typically improves Q-Scores by ~15%. Recommended for most consultations.
Web Search
When enabled, a search strategist agent analyzes your prompt and decides what to search for (your raw prompt never hits a search engine directly). Results are formatted and given to agents as context.
Search behavior follows the Intellinomics Research Doctrine:
- Thrifty: No search (saves money)
- Balanced: 2 targeted queries, single pass
- Unleashed: Up to 4 queries with a refinement round — the agent reviews initial results, identifies gaps, and searches again
Smart Recommendations
As you type, the recommender analyzes your prompt and suggests the optimal mission, forge mode, depth, Intellinomics mode, mediator, and agent team. A recommendation panel appears with "Apply All" or individual apply buttons.
The diff view shows what would change from your current settings — "was: X" badges for differences, "current" for matches.
Projects
Projects are persistent workspaces. Each project stores:
- Sessions — past consultations with full reports
- Instructions — context that every agent sees for this project
- Memory — persistent knowledge that carries across sessions
- Knowledge Base — structured entries agents reference
- Files — attached files for agent context
- Settings — saved agent config, depth, mode, etc.
The last 5 sessions are automatically provided to agents for context continuity.
Prompt Templates
Save reusable prompts with recommended settings. Templates can include variable placeholders like {{project_name}} that get filled in when used. Templates are project-scoped or global.
Export & Reports
After convergence, the full report expands to fill the screen. You can:
- Export to DOCX or PDF via the Export button
- Browse past reports via the Past Reports picker (up to 20 per project)
- Execute the plan — opens Terminal.app with Claude CLI pointed at the mediator output
- Refine — send a follow-up prompt to iterate on the report
Metacognition
Parallaxed monitors its own quality through four self-awareness systems:
- Intent Detector — auto-classifies your challenge into the best forge mode (~$0.001 per call). Override its suggestion anytime.
- Session Doctor — catches agent failures in real-time and determines recovery: retry, fall back to a different provider, skip the agent, or surface a clean error.
- Quality Reviewer — scores every synthesis 1-10 with quality flags. If the score is too low, silently re-runs the mediator with improvement instructions.
- System Reflector — periodic analysis of session patterns, provider health, and quality trends.
Missions
Missions are pre-configured agent teams optimized for specific tasks. Select a mission and the agent roles, providers, and mediator format are automatically set:
| Mission | Agents | Best For |
|---|---|---|
| Build | Architect, Dev, Research, Ops | Implementation plans |
| Code Review | Code Reviewer, Security, Performance, QA | Code audit |
| Research | Research x3, Critic | Deep analysis |
| Legal / Patent | Patent Agent x2, Compliance, Research | IP analysis |
| Brainstorm | Brainstorm x3, Creative | Idea generation |
| Debug | Dev x2, QA | Bug hunting |
| Architecture | Architect, Systems, Critic, Cloud | System design |
| Security Audit | Security x2, Compliance, Critic | Vulnerability analysis |
| Product Strategy | Product, UX, Business, Cost | Product decisions |
| Quick Fix | Brain Cramp x3 | Fast answers |
Cost Control
AI API calls cost money. Parallaxed manages this through:
- Intellinomics modes — Thrifty uses only free/cheap providers
- Monthly cost caps — each tier has a spending limit (Free: $2/mo, Pro: $25/mo, BYOK: unlimited)
- Per-session cost display — see the cost of each agent call and the session total in real-time
- Provider Fitness Matrix — automatically assigns the most cost-effective provider to each role
Voice Input
Click the microphone button next to the prompt bar to dictate your challenge. Uses your browser's Web Speech API. The mic stays active until you click it again or after 2 minutes of silence.
Keyboard Shortcuts
| Shortcut | Action |
|---|---|
| Ctrl+Enter | Start forge (CONVERGE) |
| Escape | Cancel active forge |
Pricing Tiers
| Feature | Free | Pro ($39/mo) | BYOK ($24/mo) |
|---|---|---|---|
| Sessions/day | 5 | Unlimited | Unlimited |
| Monthly cost cap | $2 | $25 | Unlimited |
| Providers | Budget only | All 14 | All (your keys) |
| Intellinomics | Thrifty | All 3 modes | All 3 modes |
| Web Search | No | Yes | Yes |
| Export | No | DOCX + PDF | DOCX + PDF |
| API Access | No | Yes | Yes |
FAQ
Why do some agents fail?
AI providers have rate limits, outages, and context window limits. Parallaxed handles this automatically: it retries, falls back to alternative providers, or gracefully skips the agent. You'll see status messages in the activity log. The Session Doctor ensures your consultation always completes.
What does the Q-Score actually measure?
It's a computational score (no LLM calls) that analyzes the text output. Divergence counts vocabulary differences between agents. Depth measures response length and structural complexity. Coverage checks topic breadth. Resolution evaluates the mediator's synthesis quality. Actionability looks for concrete next steps. It's not perfect, but it's a useful signal.
Can I use my own API keys?
Yes — the BYOK (Bring Your Own Keys) tier lets you use your own API keys for all providers. Your keys are encrypted and never leave your account. You pay the providers directly at their rates.
How is this different from just using ChatGPT?
ChatGPT gives you one perspective from one model. Parallaxed gives you 3-6 perspectives from different models that actively challenge each other. The structured disagreement surfaces blind spots that any single AI would miss. The Q-Score ensures quality. The mediator synthesizes everything into an actionable deliverable, not a conversation.
What data do you store?
Your prompts, agent outputs, and mediator synthesis are stored in your project. We never train on your data. API calls go directly to each provider (Anthropic, OpenAI, Google, etc.) under their respective privacy policies. You can delete any session or project at any time.