You can use AI to build apps without knowing how to code, but what’s actually hard is rarely generating a few screens—it’s spelling out your requirements, setting up a maintainable project, and dealing with environments, dependencies, signing, testing, and app store submission. This video documents Ada’s real workflow as he moved from Manus, Antigravity, and Android Studio to Claude Code and Codex Agent, and also covers the gotchas he hit with APK, AAB, Keystore, Play Console, and later bug fixes.
Video chapters
- 00:00 Experience and Results of Developing Apps with AI
- 00:56 Restrictions and Experience Notes Before Starting
- 01:16 First, clearly write out the requirements and the plan.
- 03:16 Phase 1: Building an app with Manus
- 04:34 Manus App Development and APK, AAB Limitations
- 05:41 Phase 2: Antigravity and Android Studio
- 07:22 Export source code, signature, and AAB packaging
- 08:31 Play Store returns and maintenance costs
- 09:32 Stage 3: Hand over to Agent to complete the process
- 11:28 From plan document to testable APK
- 12:21 Using Codex Computer Use to assist with publishing
- 14:14 Fixed Gemini Translate app and recompiled
- 15:07 Maintenance process after users report a bug
- 15:42 Source code, Keystore, version, and privacy policy
- 16:27 Where an Agent’s true value lies
- 17:30 How to Choose AI Tools and Models
- 19:01 Codex, Claude Code, and Usage Cost
- 20:29 Why aren’t we making an iOS version for now
- 21:30 Vibe Coding Reflections and Summary
Don’t rush to have AI write code: requirements and planning are the starting point.
Simply telling AI to “make me an app like a certain service” usually only produces a half-finished product that looks similar but doesn’t actually work. The video recommends first explaining the usage context, core features, input/output, and reference cases, then asking AI to expand on the requirements and organize them into a plan. Once you move into implementation, have AI develop item by item according to the plan and check each step along the way, so you don’t end up dealing with a large number of interconnected bugs all at once.
This preparation work isn’t just for people who don’t know how to code. When requirements are unclear, even a very capable model can easily lead to repeated rework and wasted tokens and time. Clearly explaining “whose problem you’re solving and what that problem is” is often more important than the first block of code.
Phase 1: Manus can quickly create prototypes, but launching is another matter.
Ada first used Manus to create Web UI and app prototypes. It can plan based on requirements and produce testable results, lowering the barrier to first-time app development. However, being able to output an APK doesn’t mean it already meets the requirements for store release; formally submitting to Google Play also involves releasable AAB, signing, versioning, compatibility, and store policies.
Although the Gemini real-time translation app shown in the video runs on a phone, moving the original project to a formal Android development environment still runs into issues with reminder features, packaging formats, and compatibility. This also reflects a common gap in Vibe Coding: demonstration prototypes can be built quickly, but a product that is maintainable, updatable, and able to pass review requires a more complete engineering process.
Phase 2: Antigravity with Android Studio
In order to test and adjust an Android project using the emulator and generate official build files, the workflow shifts to Antigravity and Android Studio. Android Studio provides a standard build environment, device emulators, and APK/AAB packaging, but if any incompatibility exists in the model, SDK, Gradle, dependencies, or original project structure, it may result in a situation where it runs on the emulator but fails on real devices or in the store.
Moving source code exported from other tools into Android Studio often causes more problems than developing from the start with the same standard environment. The video spent a lot of time dealing with signing, AAB, and Play Store rejections, and also explains why publishing and ongoing maintenance are the truly labor-intensive parts of app development.
Phase 3: Let agents like Claude Code and Codex take over the entire workflow
When AI agents can directly handle project files, run builds, read error messages, and iteratively fix issues, the workflow is no longer just pasting a piece of code into a chatbot. The video shows first asking the Agent to generate a plan, then create the app, package a test build, and then hand it over to Codex to use Computer Use to help operate the Google Play backend and complete the required fields.
This isn’t about bypassing Play Console or review rules, but rather letting the Agent operate the interfaces and workflows that would normally require manual completion. Developers still need to confirm the account, app data, permissions, store listing, and final submission content, and are responsible for the release outcome.
Data that must be kept during maintenance
Even after an app is published, you’ll still encounter error reports, device differences, and new version requirements. The video demonstrates handing screenshots and error information to Codex for analysis, then generating a fixed APK for testing. If you want AI to assist with long-term maintenance, you should at least properly preserve the source code, Keystore, version numbers, privacy policy, store backend data, and build environment documentation. Losing the signing key or project context can make future updates more troublesome than starting over.
How do you choose an AI tool? First, see if it can handle your entire workflow.
Different tools each have their own costs, quotas, model capabilities, and operation modes. Prototype services are suitable for quickly validating ideas, while professional coding agents are better suited for cross-file modifications, executing commands, handling dependencies, and iterative testing. The video also compares the user experience of Codex, Claude Code, Manus, and local models. When choosing, don’t just chase single-shot generation results; instead, consider whether it can reliably continue projects, fix errors, and fit into your existing environment.
Additionally, platform costs can affect whether an app is worth maintaining. Android and iOS have different developer accounts, listing rules, and ongoing fees, so store eligibility, testing, updates, and long-term maintenance costs should all be factored into the evaluation before building an app.
The focus of Vibe Coding is not “not needing to understand,” but rather completing validation faster.
AI has certainly made it easier for non-professional developers to turn ideas into working apps, but requirement analysis, testing, account security, store guidelines, and final acceptance still can’t be fully delegated. The most practical approach is to start by validating the workflow with a small set of features, keep recoverable source code and signing credentials, and then expand gradually—rather than asking AI to build an entire large product all at once from the start.
Source: KOCPC Chinese