Introduction: A Symphony in the Making
When we set out to build Last Apple Brain —a multi-tenant marketing automation system spanning three servers (Strategy, Orchestration, Action)—we knew it would be complex. Data synchronization with Mautic, advanced role-based security, vector-based AI operations with Weaviate, and dynamic agent execution via LangChain? That’s a lot of moving parts.
Yet the real challenge wasn’t just the system architecture. It was extracting order from tens of thousands of words across multiple chat logs—rich in ideas but scattered in focus. We realized each conversation held fragments of our final “score,” but we needed a structured process to unify them. That’s when we enlisted multiple AI models —Claude, Gemini, NotebookLM, Perplexity, ChatGPT—to help us identify blind spots, refine prompts, and systematically document our evolving solution.
Why Multiple AIs?
In the beginning, we tried using a single AI for everything: summarizing logs, answering design questions, and generating code snippets. But as the design matured, we noticed gaps:
- Technical : One AI might overlook the intricacies of CRM data flows or concurrency issues in orchestration.
- Architectural : Another AI might produce generic prompts on scaling, missing our multi-tenant security needs.
- Documentation : We also needed a partner to reflect historical chats, reminding us of decisions we’d half-forgotten.
The solution? Treat each AI model like a different section in an orchestra. We assigned distinct roles and leveraged each model’s unique “timbre”:
- Claude – The “Architectural Critic”: specialized in gap analysis, highlighting potential conflicts and reaffirming best practices around system architecture.
- Gemini – The “Prompt Crafter”: iterated on our queries, ensuring each question was precise and relevant.
- NotebookLM – The “Archivist”: combed through thousands of words from prior chats, surfacing crucial decisions or references we might have forgotten.
- Perplexity – The “Provocateur”: raised tough questions on non-functional requirements—scalability, security, data flow constraints.
- ChatGPT – The “Versatile Conductor”: bridged ideas between other models, offering summaries or clarifications when we needed a second opinion.
With these roles in place, we orchestrated an environment where each AI played its part, building a Technical Symphony of insights that gradually shaped our final system blueprint.
Enter the C.R.A.F.T. Framework
To keep each AI “instrument” in tune, we used a structured prompt approach called C.R.A.F.T. (Context, Role, Action, Format, Target Audience). Here’s a snapshot of how it brought harmony to our multi-LLM process:
- Context
We explicitly stated the problem domain: a three-server marketing automation system, with CRM (Mautic) integration, BigQuery as data warehouse, Weaviate for AI vector ops, etc.
- Role
For each prompt, we declared the AI’s perspective. For Gemini: “You are the prompt generation expert.” For Claude: “You are the system architect.” This clarified each model’s vantage point.
- Action
We outlined the exact steps (e.g., “Analyze the chat logs in doc #3 and highlight any mention of concurrency issues that could affect the Orchestration Server.”)
- Format
We requested outputs in consistent styles (tables, bullet lists, or narrative paragraphs) so we could align them easily into a living design document.
- Target Audience
We reminded ourselves that each result should speak to human specialists—Architecture, Database, Security/DevOps, AI. The goal: produce content that real people, not just LLMs, could act upon.
The Project Reality Statement: A Living Score
Each AI iteration fed into our “Project Reality Statement”—a living reference that tracked:
- System Purpose : Tying multi-tenant marketing automation to AI-powered data enrichment.
- Server Roles : Strategy (brains & data gateway), Orchestration (task lifecycle), Action (LangChain execution).
- Data Flow : How Mautic data enters BigQuery, how tasks flow from Strategy to Orchestration, and how the Action server processes them.
- Security & DevOps: Key measures like JWT-based authentication, encryption at rest and in transit, tenant isolation strategies.
- Specialist Roles : Architecture, Database, Security/DevOps, etc., clarifying who leads each design facet.
Initially, this statement was a skeleton. But with each new chat analysis—Claude cross-validating, NotebookLM recalling past design notes, Perplexity raising performance constraints—we fleshed it out. Version 1.5 eventually captured the entire architecture plus “Critical Decisions Homework,” ensuring future expansions wouldn’t stall for lack of clarity.
Iterative Insights: Rimshots & Revelations
We like to call each “aha” moment a rimshot —the crisp strike of realization that moves us forward. A few notable ones:
- Rimshot #1 : Placement of the Task State Database. Claude insisted it belonged with the Orchestration Server for concurrency safety; a single AI earlier had suggested Strategy, but that introduced latency issues. Dual-model synergy resolved the conflict.
- Rimshot #2 : Target Audience Epiphany. Gemini’s prompt design plus ChatGPT’s summary made us realize we kept writing as if the AI was the user. We pivoted to writing for actual stakeholders—human specialists—so the final documentation read like a technical reference for architects and DevOps teams.
- Rimshot #3 : Perplexity’s Security Alarm. Midway, Perplexity flagged we never documented data encryption details or tenant-based logging. That alarm led to an entire “Security Architecture” section in the Reality Statement.
Why This Matters for Complex Projects
- Multidimensional Expertise
One AI can produce an excellent design summary. Another can highlight compliance or security pitfalls. Tapping each specialized vantage point gave us a robust outcome that any single LLM alone might miss.
- Controlled Chaos
Tens of thousands of words are overwhelming. But with iterative C.R.A.F.T. prompts, we turned them into a clear narrative —like a conductor controlling an orchestra’s many voices.
- Scalable Documentation
By storing each insight in the Project Reality Statement (complete with version history), we ensured new team members or future expansions can easily re-enter the “symphony” without missing a note.
Lessons Learned & Future Directions
- Structured Prompt Engineering Pays Dividends Sloppy prompts lead to hazy answers. Crisp, targeted prompts yield immediate clarity. The time we spent meticulously refining prompts saved us countless hours in rework.
- AI-Human Partnerships Beat AI Alone If we’d simply accepted one AI’s initial suggestions, we might have overlooked concurrency strategies, multi-tenant security details, or data sync scheduling. Cross-model collaboration was game-changing.
- Iteration is Non-Negotiable Each pass revealed a blind spot. We used those discoveries to refine the Reality Statement and push the architecture forward. That cycle repeated until we converged on a stable design.
As we move forward, we’ll keep leaning on this multi-LLM synergy—especially when we tackle more specialized tasks like DevOps pipeline design or advanced analytics. Our experience confirms that orchestrating multiple AI “instruments” is worth the extra overhead, because each round of feedback either confirms a decision or prompts a better question.
Conclusion: A Standing Invitation
Building Last Apple Brain wasn’t just about technology; it was about rewriting our approach to design. By uniting multiple AI tools with a strong framework (C.R.A.F.T.) and maintaining a living Project Reality Statement, we turned chaotic chat logs into a cohesive solution that can scale, adapt, and continuously improve.
And here’s the real kicker: these methods are universal. Whether you’re designing marketing automation or a DevOps pipeline, or anything else complicated, the iterative synergy of AI plus human expertise is a formula for success. We hope our experience sparks a new wave of “Technical Symphonies” among teams facing complex challenges—where different LLM “sections” unite under a single baton, each playing precisely the notes they do best.
Interested in adopting a multi-LLM strategy for your next project? Or just want to see how our final architecture panned out?
Request a copy of the resulting Project Reality Statement via the form below and we will send it your way.