Phase 2 / Ep 07 | Human-AI Debate and PRD Fission

4 MIN READ | UPDATED: 2026-05-15

Now that our project's exoskeleton is built (with established boundaries and the using-superpowers mechanism), it's finally time to feed your "extremely rough idea" into this machine.

Let's go back to the Core Intent One-Pager you wrote in Ep01. If we were to simply throw it at a standard chat model, it would eagerly agree: "Sure, I'll start writing the code for you right away."

That's not what we want. We want a senior product manager and architect.

1. The Exclusive Prompt to Trigger Architect Mode

You need to open your terminal session (to call your Antigravity/Agent), send it the One-Pager, and include the following explicit instructions:

"[System Build Instruction] Please read the One-Pager I've provided. Before writing any code, you must enter the state of an 'extremely demanding senior system architect.' I forbid you from blindly pleasing me. You need to challenge my One-Pager with at least 3 critical questions about business or technical blind spots (e.g., offline data conflicts, bidirectional API rate limiting). After I respond, you must automatically generate a complete 'T-Block Project Implementation PRD' and save it permanently to docs/PRD_System_Design.md."

2. The Agent's Impressive Counter-Attack

Because we have GEMINI.md as our baseline and planning-with-files as its cognitive framework, the Agent won't just start working immediately. Instead, it will "push back" like a real human colleague:

  • Agent's Question 1: "You mentioned two-way sync. If a user modifies a time block offline locally, and at the same time, modifies it on their phone using the Google client, which timestamp takes precedence during the merge? Should it be Local Wins or Remote Wins?"
  • Agent's Question 2: "Since this involves dragging and dropping event blocks, the front-end calendar component will inevitably face complex cursor-based DOM rendering issues. Would you prefer me to integrate a heavy external component (like FullCalendar), or for the sake of minimalism, should I build a layout engine from scratch using CSS Grid?"
  • Agent's Question 3: "You've ruled out multi-tenancy, but what if a user switches browsers? Will their data still be accessible? Should we enforce Google SSO as the sole identifier for matching user data?"

3. Committing Decisions to a Living Document

Once you answer the Agent's questions (e.g., "Default to Remote Wins, build a lightweight Grid layout from scratch, and enforce SSO binding"), the Agent will immediately get to work, generating a detailed PRD_System_Design.md in the docs directory.

The core engineering benefit is this: This document is no longer just a manual left to gather dust in a collaboration tool. It now sits quietly in the docs/ directory. For every subsequent coding step, whenever you trigger the feature-development pipeline, the Agent will automatically first read this sacred scripture that it wrote itself.

In the next episode, after solidifying the requirements, we will have this cyber-architect perform its specialty: a relentless and rigorous [Framework and Database Selection Stress Test].