ChatGPT-4 can generate useful code chunks — single functions, UI components, debugging hints — but it cannot build a production app on its own. It has no access to your codebase, no memory between sessions, and cannot handle integrations, store submissions, or post-launch maintenance.
Short answer: ChatGPT-4 can generate useful chunks of code, but it cannot build a real mobile or web app on its own. ChatGPT is a chatbot — you can ask it for a function, a component, a SQL query, or a debugging hint and it will give you something usable most of the time. What it cannot do is architect a full system, maintain state across hundreds of files, integrate with real payment gateways, or ship an app to the App Store. For anything more than a snippet, you need either an IDE-integrated tool like Cursor or GitHub Copilot — or a professional app development team. Here is the honest 2023 state of the art.
What ChatGPT-4 Actually Is
ChatGPT is OpenAI’s conversational interface over the GPT-4 model, launched in March 2023. In November 2023, at OpenAI DevDay, OpenAI announced GPT-4 Turbo with a 128k token context window, Custom GPTs, and the Assistants API — a significant upgrade over earlier versions, but still fundamentally a chat interface.
That framing matters because “chat” is the limiting factor. Every interaction is a round-trip: you paste code in, the model responds, you paste its response back into your project, you run it, and you come back with the next question. There is no persistent view of your codebase, no ability to run the code itself, and no memory of the last session. ChatGPT sees a single window of your project at a time, and even with the 128k context window it cannot hold an entire real app at once.
What It Can Do Well
For working developers in late 2023, ChatGPT-4 is genuinely useful — provided you scope the ask correctly. Here is what we use it for day-to-day at Advisory Apps, across our current project pipeline:
- Single-function code generation. Give it a clear signature and it will write the body correctly most of the time — especially for standard patterns like REST clients, JSON parsing, form validation, or CRUD endpoints.
- Explaining unfamiliar code. Paste a 100-line function you did not write and ask “what does this do?”. GPT-4 is better at this than any previous tool.
- Debugging help. Paste the error and the offending code. You usually get a plausible hypothesis within one or two prompts.
- Boilerplate and glue. Dockerfiles, CI config, test scaffolding, RegExes, bash one-liners, shader code — everything in the category of “I know what I want, I just do not want to type it”.
- Documentation lookup. Faster than Googling the official docs for anything well-known.
- First-draft UI screens. A simple Jetpack Compose or SwiftUI screen from a text description is a 10-minute job instead of an hour.
In short: ChatGPT-4 is a good senior-pair-programmer-in-a-chat-window for small, well-bounded problems. That is a real productivity boost. It is not an app developer.
Where It Stops Working
The failure modes start the moment your problem stops being a snippet. Every one of these is something we have seen repeatedly on real client work since GPT-4 became widely available earlier in 2023.
1. It Cannot See Your Whole Codebase
ChatGPT has no access to your project files. You have to paste code in manually, one file at a time, then translate its response back. On a real mobile app with 80–200 files, this is impractical. Tools like Cursor and GitHub Copilot X (preview) solve this by running inside the IDE with the whole repo indexed, which is why they feel qualitatively different from ChatGPT for real development.
2. It Hallucinates APIs That Do Not Exist
GPT-4 confidently invents library functions, method signatures, and import paths that look correct but do not exist. This is especially common on newer libraries, recent framework versions, or anything where the training data is thin. You catch these mistakes only by trying to compile and run the code — which, in a chat window, means another copy-paste round trip.
3. It Has No Memory Between Sessions
Every new chat starts from zero. Decisions you made together yesterday are gone. Architectural choices, naming conventions, shared types — none of it persists. For a one-hour debugging session this is fine. For a multi-month app build it is disqualifying.
4. It Does Not Know Your Actual Stack
Your project has specific versions of specific libraries, a specific state management choice, a specific backend, and a specific deployment target. ChatGPT will happily give you code for a different version, a different framework, or the “popular” pattern — which may be wrong for you. A human developer reads your existing code and conforms. ChatGPT does not read your existing code at all.
5. Integrations Break It Completely
Ask ChatGPT to wire up FPX payment gateway integration, OCPP 1.6 EV charger communication, or an Oracle-backed CRM sync. It will produce code. The code will look reasonable. The code will not work, because those systems have behaviours — rate limits, session management, signature schemes, quirky timeouts — that are not in the training data and cannot be inferred from the spec. Integration work is the single largest gap between “GPT-4 snippet” and “production mobile app”.
6. It Cannot Ship
A working app has to be signed, submitted, reviewed, published, monitored for crashes, and maintained. ChatGPT cannot log into App Store Connect, cannot run your build pipeline, cannot handle a crash report, and cannot respond to a one-star review. These are not coding tasks, and they are half of the job.
If You Want More Than Snippets — What to Use
Three realistic options in late 2023, depending on where you are on the build-or-buy spectrum:
| Option | Best For | Limitations |
|---|---|---|
| ChatGPT-4 (web) | One-off questions, snippets, debugging help | Not connected to your code, no memory, hallucinations |
| GitHub Copilot / Cursor (IDE) | Working developers writing real code | Still needs a human to architect, review, and ship |
| A professional app development team | Building a production app that needs to survive real users | Higher upfront cost, longer timeline |
For a prototype you are showing to a friend, ChatGPT-4 plus a lot of patience is enough. For a side project, Cursor or Copilot in your IDE is a better bet — the loop is tighter and the context is larger. For a business application that has to integrate with real systems, handle real money, serve real users, and survive real iOS and Android updates, neither is a substitute for people who have shipped before.
Why Not Just Use Us?
Advisory Apps has been building mobile and web apps since 2012 — 11 years and 150+ delivered projects across automotive, government, healthcare, fintech, and IoT. We integrate with FPX, DuitNow, OCPP, Oracle, and most Malaysian government APIs. We run our own products in production (MedicalMet, Petairu), so we know what actually breaks after launch, not just at launch.
We use GPT-4 ourselves, every day. It makes us meaningfully faster on the parts it is good at. It does not replace the 12 years of pattern recognition our senior engineers bring to an integration, and we expect that gap to stay wide for a long time — regardless of how clever the next model announcement is.
If you have tried to build something with ChatGPT-4 and hit the wall at integrations, architecture, or scale, that wall is real. It is the point where a chatbot stops being useful and a team starts earning its keep.
Ready to Move From Chunks to a Real App?
If you have prototyped with ChatGPT and want to turn the idea into something real, the fastest next step is to talk to a team that ships production apps for a living. Book a free consultation with Advisory Apps and we will show you the closest project from our portfolio, give you a realistic MYR range and timeline, and tell you honestly which parts you can still handle with AI tools and which parts need human engineering. GPT-4 is a good co-pilot. It is not a pilot yet.
