{
  "id": 20495147,
  "title": "Your AI-coded app is not broken, it is unfinished: how to rescue a stuck Lovable, Bolt or Replit build",
  "url": "https://urgent.news/2026/11/17/your-ai-coded-app-is-not-broken-it-is-unfinished-how-to-rescue-a",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-17T21:06:21.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/samellis/your-ai-coded-app-is-not-broken-it-is-unfinished-how-to-rescue-a-stuck-lovable-bolt-or-replit-2f8l"
  },
  "original_language": "en",
  "account": "The issue with many AI-generated apps is not that they are flawed at the start, but rather become problematic during their final stages of development. This occurs around 10% of the time, often consuming approximately half of the project's effort. This pattern is consistent with applications developed using platforms such as Lovable, Bolt, Replit, Cursor, Base44, or v0. Initially, these applications may appear finished during a demo, but real-world challenges arise.\n\nFor instance, a user might sign in via Google and get redirected to a blank page, a payment may work in testing but break in production, or the database could function correctly with 400 rows but fail with more data. The generated code often produces an appealing front end but lacks essential features, such as account deletion. This is problematic when a user requests the removal of their data, as is required by GDPR regulations.\n\nThere are typically five states an AI app might find itself in:\n\n1. It operates solely on your machine, without any external deployment. It relies on local files, databases, and hardcoded API keys, using a preview backend. This is the most common situation, and it is also the easiest to resolve.\n2. The app has no error handling. The AI-generated code assumes that all calls will succeed, leading to stack traces, blank screens, or silent failures that write incorrect data and continue regardless.\n3. The data model is incomplete. It may use JSON blobs in place of tables or contain redundant tables that describe the same information. This works well for prototypes but becomes cumbersome when filtering, pagination, or migration is required.\n4. The authentication system is incomplete. Sessions may not expire, roles might exist in the user interface but not in the backend, password resets may go to no one, and an admin panel could be accessible by guessing an ID.\n5. The application cannot be deployed. There is no continuous integration (CI), no staging environment, no separation of environments, and no monitoring or rollback capabilities. Essentially, the app only exists as a laptop and a feeling.\n\nThe reasons behind these stalls are not inherent flaws in the AI coding agent. Instead, the agent optimizes for the visible features and produces a working screen, but it does not prioritize tasks that do not provide visible results, such as database migrations, adding retries, or checking user ownership on update routes. These aspects are crucial but do not create an obvious visible outcome, so the agent may overlook them.\n\nRescuing a stuck AI app involves several steps:\n\n- Project takeover: Assess the existing codebase, identify what exists versus what the original plan suggested, and document the differences.\n- Launch and deploy: Set up real environments, use real secrets, and establish a rollback plan. Most launches fail during the deployment phase, not the build phase.\n- Code optimization: Transform the prototype into a version that another developer can work with effectively.\n- Unsexy checklist: Implement session expiry, input validation, ownership checks on every write, a delete-account path, error states in the UI, and a method to identify issues in production when the app fails.\n\nBefore starting the rescue process, consider the following questions:\n\n- Can the app be deployed today, and can someone else access it via a URL?\n- What happens if the database disappears? What message does the user receive?\n- Can a user delete their account and all associated data? If not, this is a compliance issue, not a feature gap.\n- Who is responsible for on-call support, and how will they be alerted to any problems?\n\nIf you choose to handle the rescue yourself, follow these steps in order:\n\n1. Make the application deployable first – it should have a public URL, real secrets, and a rollback path, regardless of the app's appearance.\n2. Ensure that every failure is visible, as a failure you can see is a bug, while a failure you cannot see may lead to data loss.\n3. Refine the data model while the original creators still remember its purpose.\n4. Allocate time for polishing, animations, and additional features last, as these elements contribute to the product's overall feel but do not address the root causes of the issues.",
  "summary": "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…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "XDA Developers",
        "title": "I asked Lovable, Bolt, and Google AI Studio to build the same web app, and one is in a different league",
        "url": "https://urgent.news/2026/11/17/i-asked-lovable-bolt-and-google-ai-studio-to-build-the-same-web-app",
        "published": "2026-11-17T17:30:23.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}