TL;DR: "Vibe Coding"—the practice of casually guiding AI to build software—is trending. But without structure, it inevitably leads to "context drift" and spaghetti code. To build enterprise-grade software with AI, you need a rigid Context Architecture. At Mercury, we call this "The Codex Protocol." This treats your documentation not as a suggestion, but as the immutable Law that the AI must consult before writing a single line of code.
James here, CEO of Mercury Technology Solutions.
The internet is buzzing with "Vibe Coding." The promise is seductive: you speak casually ("vibes") to an AI, and it builds an app.
I have reviewed the popular workflows circulating in dev circles. They are brilliant for prototypes, but they have a fatal flaw for serious engineering: They assume the AI remembers the "Big Picture." It doesn't.
LLMs suffer from entropy. As a conversation gets long, the AI forgets the database schema you agreed upon 40 messages ago. It starts inventing new patterns. The "Vibe" turns into chaos.
At Mercury, we have refined this into a rigorous workflow. We don't call it Vibe Coding. We call it Context Engineering.
Here is the "Mercury Protocol" for building software with AI using a Single Source of Truth we call The Codex.
The Core Philosophy: "The Codex"
The secret isn't the model (Claude vs. GPT-5). The secret is Externalized Memory.
You must treat your project folder not just as a place for code, but as a "Brain" for the AI. We enforce a strict folder structure called /codex.
Why "Codex"? Because it implies authority. These are not notes; they are Laws. Before the AI writes a single line of code, it must consult The Codex.
The Structure
Create a folder named /codex in your root directory. It contains five immutable files:
- product-vision.md (The Why): The high-level goal. (e.g., "A 3D multiplayer dogfight game using ThreeJS").
- tech-stack.md (The How): The specific tools. (e.g., "WebSockets for networking, Rapier for physics. No React, vanilla JS only").
- architecture.md (The Map): The file structure and database schema. This is the most critical file.
- implementation-plan.md (The Path): A step-by-step list of tasks. Each step must have a verification test.
- progress.md (The Log): What is done, what is pending.
The Workflow: Architecture --> Execution --> Verification
Do not just open a chat and start coding. Follow this loop.
Phase 1: The Architect (Model: Claude Opus 4.5)
We use the smartest, slowest model to write the law.
Action: You act as the Product Manager. You dump your brain into the chat.
Prompt: "Act as a System Architect. Create the entries for product-vision.md and tech-stack.md based on my idea. Then, propose the architecture.md. Do not write code yet. These files will go into The Codex."
Outcome: You lock in the design. You explicitly forbid the AI from making architectural decisions later on.
Phase 2: The Scribe (Model: Claude Sonnet 4.5 / Codex)
We use the fast, 'agentic' model to execute the law.
The Golden Rule: Every new chat session MUST start with the AI consulting The Codex.
Prompt: "Consult /codex/@architecture.md and /codex/@implementation-plan.md. We are currently on Step 3. Implement Step 3 adhering strictly to the architecture defined in The Codex. Do not proceed to Step 4."
By forcing the AI to read the /codex every time, you eliminate "Context Drift." It doesn't matter if you are on prompt #1 or prompt #100; the AI is grounded in the same truth.
Phase 3: The Verifier (Model: Human + BrowserTools)
This is your job.
The popular guides suggest using automation, but I insist on Human Verification.
Action: The AI says "Step 3 is done."
You: Run the code. Check the logs. Does the 3D plane actually fly?
Crucial Step: If it works, you command the AI: "Update /codex/@progress.md to mark Step 3 complete. Update /codex/@architecture.md if you added new files."
Only then do you commit to Git. Git is your save point. If the AI hallucinates in Step 4, you git reset --hard and try again.
The "Antigravity" Advantage: Model Arbitrage
The original guide suggests choosing one tool. I disagree.
Use Model Arbitrage:
- Planning (The Architect): Use Opus 4.5. It thinks deeply about edge cases and writes a robust Codex.
- Coding (The Scribe): Use Sonnet 4.5. It follows instructions and is cheaper.
- Refactoring (The Librarian): Use Gemini 3 Pro. It has a massive context window to read the entire codebase at once if things get messy.
Conclusion: From "Vibe" to "Vision"
"Vibe Coding" implies you can just chill while the AI works. That’s false.
This workflow shifts your effort from Typing to Governing.
- You are no longer a bricklayer.
- You are the Keeper of the Codex.
Your job is to ensure the Codex stays clean. If the architecture.md gets messy, the code gets messy. If the implementation-plan.md is vague, the AI will get stuck.
Master the Codex, and you can build enterprise-grade software alone. Ignore it, and you’ll just have a folder full of broken scripts.
Mercury Technology Solutions: Accelerate Digitality.
