Most people pick one AI and stick with it.
I use two — and it's made me build faster, understand more, and actually finish projects instead of abandoning them halfway through. Here's my exact workflow, why it works, and how you can steal it.
The Setup
I use Antigravity — a coding environment by Google that lets you run AI models directly inside your editor. It has a Claude extension built in, which means I can switch between Gemini and Claude without leaving the app.
That's the key. Same environment, two models, completely different jobs.
No tab switching. No copy pasting between windows. Just a clean switch from planning mode to building mode inside one tool.
Step 1: Gemini for Planning
Before I write a single line of code, I ask Gemini to create a PRD — a Product Requirements Document.
I describe what I want to build in plain English. What it does, who uses it, what features it needs, what tech stack makes sense, what problems I might run into. Gemini is genuinely great at this. It thinks broadly, considers angles I hadn't thought of, and gives me a structured plan I can actually follow.
For my current project GeoAlert — a real-time global conflict tracker with an interactive world map, AI-generated conflict summaries, and a historical year filter — the Gemini PRD covered everything. Folder structure, which APIs to use (ACLED for conflict data, GDELT for news), how the React frontend and Python FastAPI backend talk to each other, a realistic six week build timeline, and potential problems I'd likely face.
That planning session took maybe 20 minutes. And it meant when I actually started coding, I wasn't guessing. I had a roadmap. Every decision had already been thought through.
The difference between starting with a plan and starting without one is massive. Without a plan you write code, hit a wall, rewrite everything, lose motivation, and abandon the project. With a plan you just follow the next step.
Step 2: Claude for Coding
Once the plan is solid, I switch to Claude inside Antigravity and give it one instruction:
"Use this plan and start building Stage 1."
Claude is better at staying consistent across a codebase. It follows the structure Gemini laid out, writes clean readable code, and — crucially — explains what it's doing if you ask it to.
This is where context engineering makes the real difference. I don't just say "build this." I add this to every coding prompt:
"Add a comment above every function explaining what it does. Add inline comments on any line that isn't obvious. After each stage, pause and ask me if I understand before continuing."
That one addition completely changes the output. Claude stops being a code generator and becomes something closer to a patient tutor who happens to write perfect code. Every function is explained. Every confusing line has a note next to it. I move forward only when I actually understand what just got built.
I learned this the hard way. My first project — a Rubik's Cube simulator in pygame — was built with a single vague prompt. The code worked perfectly. I had no idea why. I couldn't explain a single function. That's not a portfolio, that's a magic trick someone else performed.
The second time I built it with this workflow, I understood every line by the time it was done.
Why Two Models Instead of One?
Each model has a genuinely different strength and trying to force one model to do everything well is fighting against how these tools actually work.
Gemini thinks big picture. It's good at research, planning, weighing options, and structuring ideas across an entire project. Claude is precise and consistent. It's good at executing a specific plan, maintaining context across a long codebase, and explaining technical decisions clearly.
Using one model for everything is like using a hammer for every job. Sometimes you need a screwdriver.
The Full Workflow in 5 Steps
- Open Antigravity
- Ask Gemini: "Create a full technical plan for [project]. Include folder structure, tech stack, build order, and potential problems I'll face."
- Review the plan, ask follow up questions until it feels solid and complete
- Switch to Claude: "Use this plan. Start with Stage 1. Add comments on every function and explain each part as you go."
- Build stage by stage, understanding every line before moving to the next
That's the whole system. No magic. Just being intentional about which tool does which job.
The Bigger Point
AI tools are only as good as how you use them. A bad prompt to a great model gives you code you don't understand. A great prompt to any model gives you something you can actually learn from, explain, and build on.
The workflow I'm describing isn't about being lazy or cutting corners. It's about being deliberate, using planning tools to plan and coding tools to code, and making sure you genuinely understand what's being built at every stage.
That's not vibe coding. That's engineering.
I'm an IB student in Mumbai documenting my CS journey while applying to university.Follow along if you're on a similar path.
Top comments (0)