I constructed a backend-heavy SaaS app in Adorable and Claude Code, and the adaptation was once glaring

Vibe-coding has briefly grow to be the brand new productiveness hack for everybody. As an alternative of spending hours putting in place boilerplate, configuring databases, or wiring APIs, AI equipment can now generate running programs from a unmarried steered. However after seeing numerous demos of polished dashboards and touchdown pages, I sought after to check one thing nearer to genuine instrument building. So I constructed the similar backend-heavy SaaS app the use of Adorable and Claude Code to look which software in reality improves developer productiveness as soon as the undertaking turns into advanced and hard to debug.

Why I selected a backend-heavy app

Extra than simply fancy monitors

Somebody can steered an AI to construct a blank touchdown web page or a fundamental calculator. However in the true international, apps are not simply beautiful buttons; they’re advanced engines of knowledge, good judgment, and state control. I did not wish to see how those equipment care for CSS; I sought after to look how they care for friction.

I selected a backend-heavy undertaking, the only requiring advanced database schemas, multistep API authentications, and complicated server-side good judgment — to push those “high-agency” equipment to their verge of collapse. The objective was once to look which software may assume like a gadget architect and which one would get misplaced within the wires as soon as the UI was once completed. If an AI can’t organize the “mind” of the app, it’s only a fancy prototyper.

The detailed steered

Construct a production-style mini undertaking control SaaS utility very similar to Linear, ClickUp, or Asana.
That is NOT a frontend-only prototype. Focal point closely on backend structure, scalability, maintainability, relational knowledge modeling, permissions, and real-world SaaS patterns.
Tech Stack: Subsequent.js with TypeScript, Node.js API routes/server movements, PostgreSQL, Prisma ORM, JWT or session-based authentication, native or cloud-compatible document garage abstraction, realtime make stronger if conceivable, blank folder construction, production-style code group.
Authentication: consumer signup, login, logout, password hashing, consultation dealing with, forgot password glide, safe routes.
Organizations / Workspaces: customers can create organizations, invite individuals by means of electronic mail, organization-based knowledge isolation, multi-tenant structure.
Function-Based totally Get right of entry to Regulate: roles come with Proprietor, Admin, Member, Visitor. Permissions come with admins can invite/take away customers, visitors can solely view assigned duties, individuals can create/edit duties, homeowners can delete tasks.
Tasks: every group could have more than one tasks with fields together with title, description, standing, createdBy, timestamps.
Process Machine: duties will have to make stronger identify, description, standing, precedence, due date, labels/tags, assignee, attachments, subtasks, feedback, job historical past.
Process Relationships: duties belong to tasks, duties belong to organizations, duties could have more than one labels, duties could have more than one feedback, customers can watch duties.
Realtime Collaboration: reside activity updates, reside feedback, typing signs if conceivable.
Notifications: notification gadget for activity project, mentions, feedback, due dates, unread/learn states, notification middle.
Task Log / Audit Path: monitor activity created, activity up to date, standing modified, consumer invited, document uploaded. Retailer actor, motion, timestamp, metadata.
Seek & Filtering: seek through identify, filter out through standing, filter out through assignee, filter out through precedence, filter out through labels, sorting.
API Design: right kind REST structure, validation, middleware, authorization assessments, pagination, error dealing with.
Database Design: scalable relational Prisma/Postgres schema with indexes, international keys, many-to-many members of the family, comfortable deletes if wanted.
Record Uploads: add, delete, preview metadata for activity attachments.
UI Necessities: blank trendy UI with sidebar navigation, dashboard, kanban board, activity modal, job feed, responsive structure.
Engineering High quality: maintainable structure, reusable abstractions, separation of issues, blank code, sort protection, scalable backend patterns.
Developer Revel in: setting configuration, database migrations, seed scripts, setup directions, pattern demo knowledge.
Essential: Do NOT oversimplify the backend, do NOT faux options with mock knowledge, do NOT hardcode relationships. Construct this as though it had been an actual SaaS product anticipated to scale. Provide an explanation for architectural selections the place suitable and use real looking manufacturing patterns as an alternative of shortcuts.

Claude Code took its personal time

I used to be unquestionably now not anticipating this outcome

Claude Code was once noticeably slower from the beginning. It didn’t straight away generate polished monitors or rush via setup like Adorable did. As an alternative, it behaved extra like a cautious engineer putting in place an actual building setting. Prior to writing main options, it checked whether or not all of the required equipment and dependencies had been correctly put in — npm, PostgreSQL, programs, setting variables, gadget paths, and database connectivity. To start with, this felt pointless and sluggish, however later I noticed why it mattered.

Because the backend complexity larger, Claude Code changed into way more spectacular. Authentication flows, database relationships, API dealing with, and realtime options had been unusually solid. All over checking out, I ran right into a worm the place duties weren’t showing as it should be at the Kanban board, particularly when settling on “All Duties.” I merely described the problem to Claude Code, and it traced the basis purpose, mounted it, and up to date the code mechanically.

What stood out maximum was once its debugging skill. As an alternative of throwing random fixes on the drawback, it in reality reasoned via the problem. The workflow felt much less like “generate app speedy” and extra like “construct instrument correctly.” It unquestionably took extra time, however the ultimate codebase felt cleaner, more straightforward to handle, and a ways nearer to one thing I’d believe in manufacturing.

Adorable was once speedy

However now not sufficient!

Adorable was once simply the speedier software at first. Prior to it even began development the app, even though, it requested me to rethink the tech stack I had discussed within the steered and driven towards the use of its personal most well-liked setup as an alternative. Whilst confirming the stack, it additionally obviously discussed that the app would now not paintings within Adorable’s personal preview setting as it does now not correctly make stronger Subsequent.js.

As soon as building began, the velocity was once in truth spectacular. Inside mins, it generated polished monitors, routes, and fundamental capability that seemed production-ready at the floor. The onboarding enjoy was once clean.

However the cracks began appearing because the backend complexity larger. Some fundamental product selections had been unusually deficient. As an example, the duty introduction glide solely requested for a job title to start with, and I needed to open the duty once more simply so as to add main points later. Even document attachments, seek and filter out options, that have been obviously discussed in my steered, had been utterly lacking.

And in any case, as an alternative of absolutely serving to me send the app, Adorable passed me a ZIP document and left a number of setup steps, fixes, and integrations to be achieved manually.

Senior engineer vs. junior prototyper

After development the similar backend-heavy app in each equipment, the adaptation changed into beautiful transparent. Adorable is very good at velocity. It generates polished UI briefly, makes prototyping really feel easy, and is in truth spectacular for purchasing concepts on display screen speedy. For MVPs, demos, or frontend-heavy apps, it does a excellent activity.

However Claude Code felt a lot nearer to running with a real developer. It paid consideration to structure, debugging, backend balance, or even smaller engineering main points that almost all customers won’t understand straight away. Even on this experiment, all through database queries, Claude generated optimized queries, whilst Adorable steadily fetched whole rows unnecessarily the use of broader queries. It’s a small distinction, however the ones main points subject in genuine programs.

That’s in the end what separated the 2 stories for me. Adorable felt like a extremely smart prototyping assistant. Claude Code felt like a cautious senior engineer eager about development instrument that might in reality live to tell the tale past the demo degree.

OS

Home windows, macOS

Particular person pricing

Unfastened plan to be had; $17/month Professional plan

Workforce pricing

$100/month in line with particular person for the Max plan

Claude is an AI assistant and LLM advanced through Anthropic.