Hexad Implementation Guide

The Five AI Systems

Each AI in Hexad has distinct strengths. Knowing when to use each one is the key to effective orchestration.

🧠 Claude (Anthropic) — Integration & Safety

Best For:

• High-quality customer-facing content (emails, landing pages, product descriptions)
• Complex reasoning requiring nuance and ethical consideration
• Tool orchestration and multi-step workflows
• Crisis-aware content validation (won't suggest harmful approaches)
• Memory across sessions (Claude remembers context from past conversations)
• Code generation with strong safety rails

Use Claude When:

• Content will be seen by customers or the public
• The task requires understanding complex human dynamics
• You need an AI that "thinks before it speaks"
• Quality matters more than speed
• The topic is sensitive or could have ethical implications

Don't Use Claude For:

• Bulk content generation (expensive, use DeepSeek first)
• Current events after April 2024 (use Grok)
• Image analysis (use Gemini)
• When you need the absolute fastest response

🔥 GPT-4 (OpenAI) — Research & Analysis

Best For:

• Academic writing and research synthesis
• Code debugging (especially for complex errors)
• Data analysis and visualization
• Plugin integrations (Zapier, web browsing, Wolfram)
• Citation and reference management
• Mathematical and scientific computations

Use GPT-4 When:

• You need rigorous research with credible sources
• The task requires mathematical precision
• You're debugging code and need detailed error analysis
• You want academic-quality citations
• You need plugin capabilities Claude doesn't have

Don't Use GPT-4 For:

• Real-time current events (knowledge cutoff April 2023)
• Cost-sensitive bulk work (use DeepSeek)
• When you need the most safety-conscious responses (use Claude)

🚀 Grok (X.ai) — Real-Time Intelligence

Best For:

• Current events and breaking news
• Social media monitoring and trends
• Real-time competitive analysis
• Public sentiment tracking
• Live data that changes by the hour
• X/Twitter-specific insights

Use Grok When:

• You need to know what's happening RIGHT NOW
• You're monitoring competitors or industry trends
• The answer depends on very recent events
• You need social media sentiment analysis
• Time-sensitive decision-making

Don't Use Grok For:

• Historical analysis or timeless content
• Complex reasoning (not Grok's strength yet)
• Code generation (use Claude or GPT-4)
• Long-form content creation

🌟 Gemini (Google) — Multi-Modal Creation

Best For:

• Image and video analysis
• Processing massive documents (2M+ token context window)
• Code execution in real-time
• Google Workspace integration (Docs, Sheets, Drive)
• Visual content creation and editing
• Handling 50+ page PDFs in a single prompt

Use Gemini When:

• Working with images or videos
• You need to process entire codebases or huge documents
• You're working inside Google Workspace
• The task requires visual understanding
• You need the largest possible context window

Don't Use Gemini For:

• Pure text tasks where visual isn't needed
• When cost efficiency is critical (Gemini is expensive for text)
• Real-time current events (use Grok)

🔮 DeepSeek — Deep Analysis & Cost Efficiency

Best For:

• Long-context processing (huge documents, entire codebases)
• Cost-efficient batch processing (50%+ cheaper than Claude/GPT-4)
• First-pass analysis before refinement
• Codebase-wide refactoring
• Pattern recognition in large datasets
• Bulk content generation

Use DeepSeek When:

• Cost is a primary concern
• You need to process large volumes of information
• You're doing exploratory analysis (not final output)
• Budget is tight and you need AI assistance
• You need a first draft that will be refined by Claude/GPT-4

Don't Use DeepSeek For:

• Final customer-facing content (use Claude)
• Tasks requiring maximum safety validation
• Real-time current events (use Grok)
• When quality matters more than cost

Quick Reference: Which AI for Which Task?

Orchestration Patterns

How you route work between AIs determines your speed, cost, and quality. Here are the three core orchestration patterns.

Pattern 1: Cascade (Quality Focus)

When to Use Cascade:

• Customer-facing content that must be perfect
• Legal or compliance-sensitive materials
• High-stakes communications (investor updates, press releases)
• When quality is more important than speed

How It Works:

1. DeepSeek: Creates the first draft quickly and cheaply
2. Claude: Refines for quality, tone, and safety
3. GPT-4: Fact-checks and validates sources
4. You: Final approval and publication

Real Example: Product Launch Email

DeepSeek: "Write a product launch email for our new AI headshot service"
↓ (Copy output)
Claude: "Refine this email for professional tone and compelling CTA: [paste]"
↓ (Copy output)
GPT-4: "Fact-check all claims and verify pricing: [paste]"
↓ (Review)
You: Approve and send

Pattern 2: Parallel (Speed Focus)

When to Use Parallel:

• Research requiring multiple perspectives simultaneously
• Competitive analysis across different sources
• When you need diverse viewpoints fast
• Complex topics where no single AI has the full picture

How It Works:

1. Ask the same question to multiple AIs at once (different tabs)
2. Each AI provides their unique perspective or data
3. You synthesize the best insights from each
4. Make your decision based on combined intelligence

Real Example: Market Entry Strategy

Claude: "Should we expand into the European market? Pros/cons."
GPT-4: "Analyze European AI regulations and compliance requirements"
Grok: "What are current European tech trends and sentiment toward US companies?"
→ You: Synthesize all three perspectives into strategy decision

Pattern 3: Specialist (Targeted)

When to Use Specialist:

• Marketing campaigns requiring text + visuals
• Multi-format content creation
• When each step requires a fundamentally different skill
• Complex workflows with distinct phases

How It Works:

1. Route each task to the AI with the most relevant specialty
2. Pass outputs sequentially, each AI building on the last
3. Use the right tool for each specific job
4. You coordinate and make final decisions

Real Example: Social Media Campaign

Grok: "What topics are trending in AI education right now?"
↓
Claude: "Write 5 LinkedIn posts about [trending topics] for our audience"
↓
Gemini: "Create visual concepts for each post"
↓
You: Approve, schedule, and monitor performance

Choosing Your Pattern

Advanced: Combining Patterns

The most sophisticated Hexad workflows combine multiple patterns:

Example: Product Documentation Suite

Phase 1 (Parallel): 
  Claude: "User guide outline"
  GPT-4: "Technical reference outline"
  DeepSeek: "FAQ outline"

Phase 2 (Specialist):
  DeepSeek: Draft all three documents (bulk content)
  Claude: Refine user guide (customer-facing)
  GPT-4: Refine technical reference (accuracy)
  Claude: Refine FAQ (tone + helpfulness)

Phase 3 (Cascade):
  Gemini: Create visuals for user guide
  GPT-4: Fact-check all technical claims
  You: Final review and publish

This workflow uses all three patterns strategically based on what each phase requires.

Cost Optimization

Running five AIs doesn't mean paying for five times the work. Smart orchestration dramatically reduces costs while increasing output quality.

The Cost Reality

Total: ~$90-130/month for unlimited collaborative intelligence

Cost Optimization Strategies

1. Use DeepSeek for First Drafts

Instead of asking Claude or GPT-4 to write from scratch (expensive), use DeepSeek to create the initial draft, then refine with premium AIs.

2. Batch Similar Tasks

Group similar work together and process in bulk with DeepSeek, then selectively refine what needs polish.

DeepSeek: "Create 50 social media posts about [topic]"
↓ (Review batch)
Claude: "Refine these 5 best posts for immediate use"
↓ (Use immediately)
Rest: Keep as drafts for later quick editing

3. Route by Priority, Not Habit

Don't default to Claude for everything. Ask yourself: "Does this NEED Claude quality or can DeepSeek handle it?"

4. Parallel > Sequential for Research

Running 3 AIs in parallel takes the same time as running 1, but gives you 3x the perspectives.

Budget Allocation Framework

If you need to allocate limited AI budget across different business needs:

When Budget is Tight

If you can only afford 2-3 AIs, here's the priority order:

Tier 1 (Essential): $40-50/month

• Claude Pro - Your quality anchor for customer work
• DeepSeek API - Your cost-efficient workhorse

Tier 2 (High Value): +$20/month

• ChatGPT Plus - Research and analysis depth

Tier 3 (Complete System): +$36/month

• Grok - Real-time intelligence
• Gemini - Multimodal capabilities

Measuring Your Cost Efficiency

Track these metrics monthly:

• Cost per project: How much AI spend per completed deliverable?
• Time saved: How many hours did Hexad save vs doing it alone?
• Quality improvements: Did orchestration improve output quality?
• Revenue impact: Did better/faster content drive more sales?

If Hexad isn't saving you 10x its cost in time/quality/revenue, you're not using it right.

Real Workflows

Here are actual production workflows used to build and run real businesses using Hexad.

Workflow 1: AI Headshot Service Launch

Goal: Launch ChispaModels (AI headshot generation service) from concept to revenue in 2 weeks

Phase 1: Market Research (Day 1-2)

Grok: "What are people saying about AI headshots on X/Twitter right now?"
→ Identifies pain points: expensive photographers, inconsistent quality

GPT-4: "Research competitor pricing and features for AI headshot services"
→ Finds pricing range: $29-$99, identifies feature gaps

Claude: "Based on these insights, what's our positioning?"
→ Recommends: $39 price point, focus on professional quality + speed

Phase 2: Product Development (Day 3-7)

DeepSeek: "Write technical spec for LoRA training pipeline"
→ Cheap first draft of architecture

GPT-4: "Review this spec for technical accuracy and optimize"
→ Validates approach, suggests improvements

Claude Code: "Build the MVP from this spec"
→ Writes production code with safety rails

Phase 3: Marketing Materials (Day 8-10)

DeepSeek: "Write 20 variations of landing page copy"
→ Bulk content generation

Claude: "Refine the best 3 versions for conversion"
→ Quality polish for customer-facing content

Gemini: "Create visual concepts for the landing page"
→ Image suggestions and layouts

You: A/B test and select winner

Phase 4: Launch (Day 11-14)

Claude: "Write launch announcement for email list"
→ High-quality customer communication

Grok: "Monitor social media reaction and sentiment"
→ Real-time feedback tracking

GPT-4: "Analyze conversion data and suggest optimizations"
→ Data-driven improvements

Workflow 2: Weekly Content Creation

Goal: Create a week's worth of educational content (blog posts, social media, newsletter)

Monday: Topic Research

Grok: "What's trending in our industry this week?"
→ Gets 5 current trending topics

Parallel (Claude + GPT-4 + DeepSeek): "Which of these topics should we write about?"
→ Each AI gives perspective, you choose top 3

Tuesday: Long-Form Content

GPT-4: "Research [topic 1] - give me sources and outline"
→ Academic rigor and citations

DeepSeek: "Write 2000-word blog post from this outline"
→ Cheap bulk content

Claude: "Refine this post for our brand voice and add compelling hooks"
→ Quality polish for publication

Wednesday: Social Media

DeepSeek: "Create 20 social posts from blog content"
→ Batch generation

Claude: "Select and refine the best 7 for this week"
→ Quality curation

Gemini: "Suggest images for each post"
→ Visual recommendations

Thursday: Newsletter

Claude: "Write engaging newsletter featuring this week's content"
→ Customer-facing communication

GPT-4: "Fact-check all claims"
→ Accuracy validation

Friday: Analysis & Optimization

Grok: "How are our posts performing? What's the sentiment?"
→ Real-time feedback

GPT-4: "Analyze performance data and suggest next week's topics"
→ Data-driven planning

Workflow 3: Customer Support Crisis

Scenario: Service outage affects 50+ customers, need immediate response

Crisis Mode Activation

You: "🚨 EMERGENCY - service down, customers affected"

Grok: "Monitor social media for customer complaints and sentiment"
→ Real-time tracking of public reaction

Claude: "Write immediate status update email (honest, empathetic, clear)"
→ Customer communication

GPT-4: "Draft technical post-mortem and prevention plan"
→ Technical documentation

DeepSeek: "Analyze error logs and identify root cause"
→ Technical analysis

Claude: "Write personalized apology emails to affected customers"
→ Relationship management

Resolution Timeline

• 0-15 min: Claude sends status update to all customers
• 15-60 min: DeepSeek identifies issue, GPT-4 documents, you fix
• 60-120 min: Grok confirms sentiment improving, Claude sends resolution email
• Next day: GPT-4 completes post-mortem, Claude sends follow-up

Workflow 4: Product Documentation Suite

Goal: Create comprehensive documentation (user guide, technical reference, FAQ, video scripts)

Step 1: Parallel Outline Generation

Claude: "Create user guide outline"
GPT-4: "Create technical reference outline"
DeepSeek: "Create FAQ outline"
→ All three run simultaneously

Step 2: Bulk Content Generation

DeepSeek: "Write all three documents from these outlines"
→ Cost-efficient first drafts

Step 3: Specialized Refinement

Claude: "Refine user guide for clarity and helpfulness"
→ Customer-facing quality

GPT-4: "Refine technical reference for accuracy and completeness"
→ Technical precision

Claude: "Refine FAQ for common questions and friendly tone"
→ Support quality

Step 4: Visual Enhancement

Gemini: "Create diagrams and screenshots for user guide"
→ Visual aids

Claude: "Write video scripts for tutorial series"
→ Multimedia content

Handoff Protocols

Clean handoffs between AIs are critical for velocity. Sloppy handoffs waste time and create confusion.

The Handoff Template

When passing work between AIs, use this standardized format:

## From: [AI name]
## Summary: [What was done in 1-2 sentences]
## Output: [The actual content/work]
## Open Questions: [What needs resolution or clarification]
## Next Step: [Specific action for receiving AI]

Good vs Bad Handoffs

DeepSeek → Claude:

"Here's the content"

[giant paste]

DeepSeek → Claude:

From: DeepSeek
Summary: Created first draft of product launch email
Output: [content]
Open Questions: Should we emphasize speed or quality?
Next: Refine for brand voice and add compelling CTA

AI-to-AI Handoff Examples

Example 1: Research → Writing

GPT-4 → Claude:

From: GPT-4
Summary: Completed competitive analysis of 5 AI headshot services
Output:
- Competitor A: $49, 100 photos, 24hr turnaround
- Competitor B: $29, 50 photos, 48hr turnaround
- etc...

Open Questions: Which differentiator should we emphasize most?
Next: Write positioning statement based on this analysis

Example 2: Draft → Refinement

DeepSeek → Claude:

From: DeepSeek
Summary: Generated 10 blog post drafts on AI safety
Output: [10 drafts pasted]

Open Questions: None - all drafts complete
Next: Select best 2 and refine for publication quality

Example 3: Analysis → Visualization

GPT-4 → Gemini:

From: GPT-4
Summary: Analyzed user behavior data, found 3 key insights
Output:
1. 80% of users drop off at pricing page
2. Mobile users convert 2x better than desktop
3. Users from social ads stay 3x longer

Open Questions: None
Next: Create visual dashboard showing these insights

You-to-AI Handoffs

When you're giving work to an AI, be equally clear:

You → Claude:

Task: Write investor update email
Context: We hit $10K MRR this month, 2x growth from last month
Tone: Professional but excited
Length: 200-300 words
Key points to include:
- MRR milestone
- User testimonial (see attached)
- Next month's roadmap preview

Success looks like: Email I can send with minimal editing

Escalation Triggers

Sometimes you need to switch AIs mid-task. Here are the triggers:

Escalation Handoff Format

## Escalation from [AI 1] to [AI 2]
## Reason: [Why switching]
## What [AI 1] tried: [Brief summary]
## What didn't work: [Specific blockers]
## What [AI 2] should try: [Your suggested approach]
## Context: [Any relevant background]

Real Escalation Example:

GPT-4 → DeepSeek Escalation:

Escalation from GPT-4 to DeepSeek
Reason: Need to process entire 500-page codebase for refactoring
What GPT-4 tried: Analyzed sample files, provided recommendations
What didn't work: Can't hold full codebase in context
What DeepSeek should try: Scan full codebase, identify all instances of deprecated API
Context: Preparing for Q2 migration, deadline is end of month

Parallel Handoffs

When you're running AIs in parallel, you're handing off to YOURSELF for synthesis:

Claude || GPT-4 || Grok → You:

Your synthesis prompt to yourself:
- What did each AI uniquely contribute?
- Where do they agree/disagree?
- Which perspective is most actionable?
- What's my final decision based on all inputs?

Emergency Protocols

Hexad includes specific protocols for when things go wrong. Crises require different orchestration patterns than routine work.

🚨 Crisis Mode

Activation Triggers:

• Service outage affecting customers
• Security breach or data concern
• Customer safety issue
• Legal/compliance emergency
• Reputational crisis (bad press, viral negative feedback)
• You explicitly say: "EMERGENCY", "CRISIS", or "🚨"

When Crisis Mode Activates:

Crisis Mode AI Assignments:

Crisis Communication Template (Claude):

Subject: [Clear, honest description of issue]

What happened:
- [Specific facts only, no speculation]

Who is affected:
- [Exact scope - don't minimize or exaggerate]

What we're doing:
- [Concrete actions being taken RIGHT NOW]

What you should do:
- [Clear, actionable steps for affected users]

Timeline:
- [When they can expect next update]

Our commitment:
- [How you'll prevent this going forward]

💰 Revenue Emergency

Activation Triggers:

• Revenue-critical service down (e.g., ChispaModels payment processing fails)
• Major customer churning
• Competitor launches game-changing feature
• Viral negative review affecting sales

Revenue Emergency Protocol:

Immediate (0-30 min):
You: Assess business impact and make triage decision
Claude: Draft customer communication (if needed)
Grok: Monitor competitor moves and market sentiment

Short-term (30 min - 4 hours):
DeepSeek: Technical analysis of root cause
GPT-4: Document timeline and business impact
Claude: Coordinate customer retention efforts

Resolution (4 hours - 24 hours):
You: Implement fix + business continuity plan
GPT-4: Post-mortem analysis
Claude: Customer follow-up communication
Grok: Monitor sentiment recovery

🛑 When the Anchor Says "Stop"

Sometimes you (the human anchor) need everything to just STOP. This protocol ensures that happens cleanly.

Stop Protocol:

1. All AIs stop immediately - No completing current thoughts, no "but let me just..."
2. No explanations - Don't ask why, don't justify what was being done
3. Ask: "What's the actual blocker?" - One question only
4. Wait for direction - Don't suggest next steps, don't try to help yet

Why This Matters:

When you're frustrated, overwhelmed, or stuck, the LAST thing you need is more AI output. You need space to think. The Stop protocol creates that space.

⚠️ Ethical/Safety Concerns

Activation Triggers:

• Request involves potential harm to users
• Legal gray area or compliance concern
• Deceptive marketing or business practice
• Privacy violation possibility
• Any "this feels wrong" instinct from you

Safety Protocol:

Step 1: Immediate Pause
All AIs stop executing on the task

Step 2: Claude Safety Check
Claude: "I need to flag potential concerns with this approach..."
Claude provides specific safety/ethical considerations

Step 3: Human Decision
You assess whether to:
a) Proceed with modifications
b) Choose different approach entirely
c) Abandon task

Step 4: Document Decision
GPT-4 documents your reasoning for future reference

Emergency Contact Tree

Know who to escalate to for different emergencies:

Post-Crisis Review

After every emergency, conduct this review within 48 hours:

GPT-4: "Create timeline of what happened"
Claude: "Analyze communication effectiveness"
Grok: "Report on public sentiment arc"
You: "What do we change in our protocols?"

Update your Hexad emergency protocols based on lessons learned.