DEV Community

Sam Ellis
Sam Ellis

Posted on

Your AI-coded app is not broken, it is unfinished: how to rescue a stuck Lovable, Bolt or Replit build

Most AI-coded apps do not fail in the first week. They fail in the last 10%, which is usually 50% of the work.

That is the pattern I keep seeing: Lovable, Bolt, Replit, Cursor, Base44 or v0 produces something that looks finished in a demo, and then something real happens. A user signs up with Google and the OAuth redirect drops them on a blank page. A payment succeeds in the test environment and breaks in production. The database works until there are 400 rows. The app has a beautiful front end and no way to delete a user account, which is fine right up until a GDPR request arrives.

The coding agent did what it was asked. The product was never finished.

The five states a stuck AI app is usually in

1. It runs on your machine and nowhere else. Secrets in a .env file, a local database, a hardcoded API key, a backend that only exists inside a preview. This is the most common state by a distance, and it is also the cheapest to fix.

2. It has no error handling. The generated code assumes every call succeeds. You get a stack trace, a blank screen, or a silent failure that writes a bad row and moves on. Nobody tested it because there was no test to run.

3. The data model is a sketch. A JSON blob where a table should be, or three tables that describe the same thing. Fine for a prototype, painful the moment you need to filter, paginate or migrate.

4. The auth is half-wired. Sessions that do not expire, roles that exist in the UI but not in the backend, a password reset that sends to no one, an admin panel reachable by guessing an ID.

5. Nobody can deploy it. No CI, no staging, no environment separation, no monitoring, no way to roll back. The app exists only as a laptop and a feeling.

Why it stalls

An AI coding agent optimises for the visible thing. You asked for a working app, so it produced a working screen. It has no incentive to write the migration, add the retry, or check the ownership on the update route, because those produce no visible result. It also cannot tell you what it left out, because it does not have a model of your intent - only of the request.

So the last 10% is not a bug in the tool. It is the part of building software that was never a "build a feature" task: verification, error paths, deployment, and the boring correctness work that has no screenshot.

What a rescue actually involves

If you want an outside pair of hands, https://www.appstuck.com/ does this as a service: they take over an existing codebase, map what is there, work out what is missing, and take it to a launch. The useful part is not the code - it is the checklist. A rescue usually runs through:

  • Project takeover. Reading the codebase as it is, not as it was intended, and writing down what actually exists versus what the plan says exists.
  • Launch and deploy. Real environments, real secrets, real rollback. Most "launches" that fail do so on the deploy step, not the build step.
  • Code optimisation. Turning the prototype into something a second developer can work in.
  • The unglamorous checklist. Session expiry, input validation, ownership checks on every write, a delete-account path, error states in the UI, and a way to see what happened in production when it breaks.

The questions worth asking before you start

  • Can I deploy this today, to a URL somebody else can open? If no, you have a prototype, not a product.
  • If the database disappears, what does the user see? If the answer is a stack trace, that is your next afternoon.
  • Can I delete my own account and all my data? If not, you have a compliance problem, not a feature gap.
  • Who is on call, and how do they know? The alternative is that you find out from a user.

Where to start if you are doing it yourself

Work through the states in order. Get it deployable first - a public URL with real secrets and a rollback path, even if the app is ugly. Then make every failure visible, because a failure you can see is a bug and a failure you cannot see is a data-loss incident. Then harden the data model while you still remember what it means. Leave the polish, the animations and the extra settings for last, and accept that they are the part that makes people say the product feels good.

Most of the time, the honest answer to "is my AI-coded app finished?" is no, and the gap is smaller than it feels. It is about half a week of focused work for a small app, and considerably more for a large one. The question is only whether you want to spend it yourself.

Disclosure: I work with the team behind AppStuck.

Top comments (0)