Prompting Opus 4.8 or GPT-5.6 to scaffold a multi-page Next.js 16 app burns hundreds of dollars in API tokens and hours of agent baby-sitting. Here is the math behind why cloning beats prompting.
Last month I decided to test a simple hypothesis: with top-tier models like Claude Opus 4.8 and GPT-5.6 running inside CLI agents, is the SaaS boilerplate officially dead?
I opened a fresh directory, booted up an agentic loop, and asked it to scaffold a production-ready Next.js 16 app with Better-Auth, Stripe subscriptions, Tailwind v4, and Drizzle ORM.
Three hours and a $180 Anthropic invoice later, I had my answer.
When people say "AI coding is basically free," they are usually talking about generating a single UI component or a standalone utility function. That costs pennies.
That math completely falls apart the second you ask an agent to build a multi-page architecture across 30+ interconnected files.
Agentic workflows don't just write code—they inspect directory trees, run terminal checks, read type definitions, and edit files across multiple passes. Every time the agent makes a tool call, runs npm run build
, or checks why a Server Component failed to re-render, it re-sends the entire repository context back into the model.
Add thousands of reasoning tokens per pass, and a single multi-turn session to wire up auth, session cookies, and Stripe webhooks easily burns through $150 to $300+ in raw API credits.
You are effectively paying AI providers hundreds of dollars to have an agent repeatedly re-invent standard authentication flows from scratch.
Models like Opus 4.8 and GPT-5.6 are brilliant at generating localized logic. But a production SaaS isn't a collection of isolated files—it is a tightly coupled web of dependencies.
A real foundation requires 70+ modules and 30+ database tables all speaking the exact same language:
When you try to prompt an agent into building this graph from scratch, type drift is inevitable. The database schema slightly diverges from the session middleware; the RBAC checks miss an edge case in server actions; the payment webhooks break when a workspace ID is passed as undefined
instead of a string.
Fixing these cross-module integration bugs requires constant back-and-forth prompts, burning through tokens just to get basic features to stop breaking each other.
Agents excel at building things that run on localhost:3000
via npm run dev
. They are notoriously unreliable when preparing code for production infrastructure.
Getting a project ready for production means setting up Infrastructure as Code (IaC):
DB_PROVIDER
).When you ask an agent to generate valid Helm charts or Terraform manifests alongside app code, it frequently hallucinates configuration keys or outputs outdated syntax. Debugging a failing deployment pipeline through an AI agent can cost $30 to $50 in API compute alone, purely spent fixing minor syntax errors in YAML and HCL files.
Code generated by a prompt session is a static snapshot in time. It has no lineage and no upstream repository.
Six months from now, when Next.js introduces breaking API changes, Tailwind v4 releases a major patch, or security vulnerabilities surface in your auth dependencies, your AI-generated starter becomes instant technical debt. Your only recourse is to feed the whole codebase back into an agent and hope it doesn't break your custom business logic while refactoring.
A purpose-built boilerplate provides a Git upstream remote. When core security patches, framework migrations, or new features drop, you pull them directly with git pull upstream
. You own a living codebase backed by an active maintainer, not an unmaintainable single-shot generation.
Even if API tokens were completely free, there is the time cost.
Current agents are fast, but they aren't autonomous enough for you to write one prompt and walk away while a full application builds itself. You are forced to sit at your terminal baby-sitting the process:
`cookies()`
, `headers()`
, or route params
).If you ever decide to customize the setup mid-build—for instance, removing multi-tenancy while keeping organization billing—asking an agent to refactor its own freshly written code usually results in broken imports and orphaned schema tables.
Compare that to running an interactive CLI installer like npx create-codapult
, which prompts you for feature flags upfront and instantly outputs a clean, trimmed codebase in seconds.
Top-tier models are incredible at writing complex, unique business logic that sets your product apart. Using high-reasoning compute to re-generate standard Stripe subscription lifecycles, Tailwind v4 configs, and DB schemas is a waste of capital.
You’re paying premium rates for the AI equivalent of laying a brick foundation.
This is exactly why I built ** Codapult**.
It is a clean, modular Next.js 16 SaaS starter designed to save you from prompt fatigue and context drain:
You run git clone
or use npx create-codapult
, drop in your .env
variables, and start building actual features immediately.
Save your high-tier tokens for your core product.
Codapult includes a complete, production-ready Next.js 16 stack out of the box — Better-Auth, Stripe/LemonSqueezy billing lifecycles, Tailwind v4, Drizzle ORM, IaC configs, and modular plugins — saving you hundreds of dollars in API compute and hours of agent orchestration. Explore the architecture in the documentation.