zeiki

Building Real SaaS with AI: Not Just Side Projects

December 14, 2025·zeiki

1️⃣

idea

Over these past few months grinding on AI SaaS, one thing completely shifted my perspective:

Using AI for "vibe coding" on side projects is fundamentally different from using AI to build a SaaS that actually makes money and can be maintained long-term.

2️⃣

Most people treat GPT-5 / Claude / Codex like a smarter Copilot.

They churn out a bunch of "good enough to run" code, then leave the technical debt for their future selves to deal with.

The result? A pile of demos, but zero products that actually generate revenue.

3️⃣

My core shift now:

I've stopped chasing "ship a product in one night."

Instead, I use LLMs as a "remote technical co-founder" to build SaaS like a real company, step by step—not as a toy project.

4️⃣

Model Division of Labor

GPT-5.1 HIGH : Product Manager + Architect (defines roadmap, EPICs, tech stack, risk management)

Claude Opus 4.5 / GPT-5-Codex : Handles implementation for each development PHASE—writes code + tests, no shortcuts.

5️⃣

My SaaS Development Flow

  1. First, have GPT-5 write the complete product EPIC: business objectives, core metrics, pricing model, MVP scope

  2. Break it down into executable PHASEs by month/week (backend, frontend, payments, permissions, growth loop)

  3. For each PHASE, generate a "super prompt for the code agent"

workflow

6️⃣

Then this prompt goes to Claude / Codex:

It only handles code and tests for this specific phase—no "adding features on the fly."

Once that phase is done, go back to GPT-5 for code review, focusing on "does this help the business?" not just "does the code look pretty?"

7️⃣

With this approach, as a solo dev, I'm actually managing a small team:

  • LLMs handle roadmap, ROI assessment, complexity control

  • Code models do the repetitive work

  • Humans do three things: pick the market, make decisions, ship releases

8️⃣ Practical Implementation

In the terminal, I run multiple panes (Rio / WezTerm work great):

  • Left pane: Codex CLI with GPT-5.1 HIGH—the "architect window"

  • Right pane: Codex / Claude—the "execution window"

The entire workflow from Git repo, testing, to deployment runs in this closed-loop "AI team."

9️⃣

The best part about building SaaS:

When you have your AI team write code around "customer value" and "subscription retention" instead of "finish the features," you realize most of those cool features you wanted to build? You don't actually need them.

🔟

If you're currently working a day job while trying to build SaaS with AI:

Instead of starting yet another side project, refactor your AI workflow first.

Shift from "let's hack something together today" to "build myself a product that can last 1-2 years and generate consistent revenue."