We Benchmarked Claude Code 36 Times Building SaaS Backends — Here's What Broke

Choleski — Tue, 28 Jul 2026 05:41:36 +0000

We asked Claude Code (opus-4-8) to build the same growing SaaS app three ways, three times each: with Fonderie, with our full skill loaded, and from scratch with no Fonderie knowledge. 36 sessions total.

The from-scratch runs:

Reinvented auth, billing, and permissions every single time (different schemas, different API shapes)
Hardcoded a JWT secret in 2 of 3 runs
Burned ~3× the tokens per turn

The Fonderie runs:

Reached for pre-built, audited bricks instead of improvising
Zero hardcoded secrets (the brick throws if env vars are missing)
~⅓ the code, ~⅓ the context-window pressure

Fonderie is an open-source backend skill for AI coding assistants. Instead of letting your LLM re-derive auth, Stripe billing, workspaces, and RBAC from scratch every session, you give it a standard mold it already knows.

One command:

npx create-fonderie-app my-saas

Then tell Claude: "Add billing." It installs @fonderie/billing, wires the routes, and moves on to your actual product.

MIT licensed. No seats, no rent. The benchmark harness and raw transcripts are in the repo.

Repo: https://github.com/fonderiejs/fonderie
Docs: https://fonderiejs.com

DEV Community: Choleski

We Benchmarked Claude Code 36 Times Building SaaS Backends — Here's What Broke