Finishing an app built with Cursor, Claude Code or Codex

You had a clear idea, the tools were good, and in a few weekends there was an app. Screens, a database, login, the first real users. Each feature came quicker than the last, until one didn't. A change in one place broke three others. The assistant started patching its own patches. Fixed bugs came back. And the feature the business needs next, the one that touches everything, won't come.

None of that is a failure of the tools. The assistant wrote what was asked, one request at a time, and nobody was holding the whole design. That's the job that was missing.

More than most rescue cases. The code is in a language and framework a developer can read: usually TypeScript with Next.js or React, or Python, with Postgres or Supabase behind it. It's in a Git repository, or can be in one this afternoon. It runs on your machine and probably on a host you chose. There's no platform holding the back end, no export to negotiate, no per-user licence.

That's the difference from a Lovable, Bolt or Bubble app. The question isn't how to get the code out. It's what the code is missing.

• A designed data model. Tables were added as features needed them. Now the same fact lives in two places and they disagree.
• Tests. The app was tested by using it. Every change is a chance to break something that worked last week, which is why the last features won't come.
• Authorisation. Login exists. The check that one customer can't see another's data often doesn't, on every route.
• Environments. One database, and it's live. Changes are tried on real data.
• Secrets. API keys in the repository, or in a file on one laptop.
• Structure. Files that grew to thousands of lines, logic in the screens, the same thing done four ways. The assistant only ever saw the part it was asked about.
• Operations. Error reporting, logging, a backup that's been restored, a deployment that doesn't depend on your machine.

Finish, nearly always. The stack is mainstream, the screens are what your users already know, and the data model is close. The work is foundations first (tests around what works, environments, authorisation, secrets, a data model straightened out), then the structure that's blocking progress reworked where it needs to be, then the features in order of value.

Rebuild when the structure has stopped accepting changes at all, and even then the existing app is the specification and the data carries across. The general keep, finish or replace reasoning applies, and the assessment makes the call from the code.

1. Repository access, and a build on a clean machine. This is where the missing secrets and the laptop-only steps show up.
2. Read it. The assessment: every route, every table, every place a rule is enforced or isn't, every outside service, and the structure that's blocking the next feature.
3. Foundations. Tests around the processes that work today, so they keep working. A test environment and a staging environment. Authorisation on every route. Secrets in a store you control. A backup that's been restored.
4. Straighten what blocks. The data model, and the parts of the structure that every new feature trips over. Not a rewrite; the minimum that makes the next feature possible.
5. Finish the features, one at a time, each priced before you commit.

We use the same tools you did, with tests, review and a design held by a person. That's the part that was missing, not the tools.

It's already in a database you own, which is the good news. It never meets a change until the change has proven itself somewhere else: tested on a copy in a test environment, then run in staging with the data reconciled, before anything connects to live. Where the data model changes, the migration is rehearsed the same way.

The assessment is from £395 + VAT, sized on a free one-hour consultation, with an exact price before you commit. For an assistant-built app it reads the code, lists what's missing and what's blocking, and prices the foundations and each remaining feature as fixed numbers. What drives the price and the payment terms each have a page.

Read it, make it safe, unblock it, finish it. Marc Allington builds with AI coding tools daily, Claude Code among them, with the tests and review around them that keep a codebase moving, so he can tell quickly what an assistant-built app has and hasn't got. Everything we finish runs on open technology in your own accounts, with full source code and ownership staying with you throughout.