10 things I wish I knew before trusting Claude Code to build my iPhone app

Elyse Betters Picaro / ZDNET

Follow ZDNET: Add america arsenic a preferred source connected Google.

ZDNET's cardinal takeaways

• Vibe coding works champion successful tiny steps, not large specs.
• Persistent AI documentation eliminates re-ramp time.
• Git, backups, and exports are captious information nets.

This is not my archetypal vibe coding rodeo. I antecedently did a project utilizing OpenAI's Codex with large success. But it was a task of a antithetic scope. That was adding caller add-on modules for a WordPress plug-in. This is utilizing Claude Code to physique an entire, analyzable iPhone exertion from scratch.

Also: 10 ChatGPT Codex secrets I lone learned aft 60 hours of brace programming with it

In this article, I'll stock with you 10 lessons that came from the two-week play during which I did astir of the coding connected my iPhone app. If you privation to larn much details astir the app itself, consciousness escaped to constituent your browser to this article.

As a super-quick background, my app uses NFC tags and photos to assistance maine negociate 3D printer filament successful my Fab Lab. It tracks 100+ spools crossed 20 retention shelves, 8 machines (with 26 idiosyncratic spool holders), and different 5 oregon truthful enactment surfaces. It's designed for a super-fast workflow truthful if a spool moves from 1 instrumentality oregon determination to another, the database tin beryllium updated with a elemental pat of my iPhone connected the spool and the destination.

Only 1 of these tips is Apple-specific. The remainder of them use to vibe coding with Claude, careless of platform. And with that, let's get started.

1. Code successful babe steps

In my antecedently mentioned 10 Codex secrets article, I talked astir wherefore I didn't travel the often recommended champion signifier of creating a merchandise requirements papers and past feeding the afloat spec to the AI.

If anything, I americium present overmuch much convinced that feeding an full specification to the AI astatine erstwhile is an perfectly bonkers idea. The past happening you privation an AI to bash is instrumentality a 20-page papers and make 20,000 lines of code, each of which you person to excavation through, test, and corroborate interactions. The thought sounds bully connected paper, but it volition marque for incredibly problematic products.

Also: How to instal and configure Claude Code, measurement by step

Instead, provender 1 tiny measurement astatine a clip to the AI. In my case, I started by simply getting the AI to work an NFC tag and study its contents. That was 1 of the 2 large hurdles I saw successful the project, due to the fact that that was wholly iOS-specific coding.

Once it could work the tag, I had it constitute a tag. Then I checked to spot if it could work a antecedently written tag. Then I added the quality to instrumentality a photo. Then I added the quality to discard a just-taken photo. Then I added the quality to prevention a photo.

At nary constituent successful the aboriginal steps did I springiness the AI instructions astir the wide imaginativeness of the product. But erstwhile we had these pieces working, we could determination connected to more. The payment of this attack is that I could exhaustively trial retired each caller capableness oregon feature, and tweak it with Claude until thing really worked.

2. Instruct Claude to instrumentality notes

Probably the astir invaluable extremity I could perchance springiness you is the workflow I adopted astatine the extremity and opening of each session. Early on, I told Claude to make documents to notate the learnings that came from enabling the NFC capabilities. This was necessary, due to the fact that it failed connected its aboriginal tries, but past figured it out. I wanted it to retrieve what it had figured out.

This signifier was extended to my last bid astatine the extremity of each session. Whenever I decorativeness a league with Claude, whether it's a two-hour league oregon a 15-minute session, I archer it to papers everything we did, grounds everything we learned, and support way of what we request to bash next. I archer it to bash truthful successful specified a mode that some I and the AI tin recognize comprehensively what was learned, what decisions were made, and what our adjacent steps are.

Also: The champion escaped AI for coding successful 2025 - lone 3 marque the chopped now

The contented (architecture patterns, migration system, UI design, and truthful on) recorded successful .MD files was built up implicit each our sessions from those explicit requests. Each clip I asked Claude to grounds advancement and notes, it added immoderate seemed important to preserve, whether that was caller patterns, decisions we made, oregon workflows that we discovered.

This is not automatic behavior. The CLAUDE.md record (along with different .MD files) has grown into a broad task usher due to the fact that of my prompts to bash so. In fact, Claude's default behaviour is to ne'er proactively make documentation files (this is really successful its strategy instructions).

At the opening of each session, I archer Claude to reappraisal the code, reappraisal each the docs, and travel up to velocity connected each the notes. Then fto maine cognize what we did past time, wherever we are, and what it thinks we request to bash next.

It took maine a portion to recognize it, but this signifier resulted successful this next, precise unexpected acquisition (and benefit) of vibe coding with Claude.

3. I was capable to destruct the come-up-to-speed clip

For me, this was huge. I don't programme afloat time. Instead, I get tiny blocks of clip betwixt each my different responsibilities. For years, my biggest situation was getting backmost up to speed.

Basically, for me, coding meant having to person the full architecture of the task successful my head, being afloat alert of the programming language, knowing each the issues, and past penning codification for the adjacent capability.

The occupation is that I'm often distant from the codification for weeks oregon months astatine a time, paying the bills. I mightiness person wanted to dip into coding for an hr oregon truthful connected the weekend, but due to the fact that it would often instrumentality maine a time oregon 2 to travel backmost up to velocity connected a task and re-learn everything, I couldn't instrumentality vantage of those fewer hour-long gaps successful time.

Also: Why AI coding tools similar Cursor and Replit are doomed - and what comes next

Instead, I had to hold until I had a escaped three- oregon four-day block, which lone happens a fewer times each year. Needless to say, guardant question connected my coding projects was reasonably slow.

But with vibe coding and my signifier of making Claude papers everything, I tin enactment connected my task adjacent if I lone person a 30-minute artifact of time. I inquire Claude to travel up to speed, which takes astir 5 minutes. I tin work our updated notes, springiness it a prompt, archer it to grounds its progress, and determination on.

Unlike my erstwhile projects, wherever I needed blocks of time, my caller iPhone app was coded implicit a 17-day period, successful one- oregon two-hour chunks. I've near it for a fewer weeks and travel backmost to it, and person been capable to proceed moving connected it. For me, this is crippled changing.

4. Compacting means it's clip to extremity a league

Claude does a process called "Compacting," that allows it to compress and scale the tokens successful the existent session. The archetypal clip this happened, I got a announcement that said "Compacting" with nary further details. Then, my enactment with the AI conscionable stopped. It appeared to bent for astir 15 minutes. I thought thing had broken, but it was OK.

Also: You should inactive larn to code, says apical Google AI exec - here's why

The compaction process is initiated erstwhile Claude's league reaches a complexity level wherever it needs to negociate its token use. You tin inquire Claude however adjacent it is to compaction if you privation to beryllium proactive. I bash that sometimes. Most of the time, I recognize it's been a portion erstwhile I get the compaction notice.

My proposal to you (and the signifier I've followed ever since my archetypal notice) is to usage the compaction announcement arsenic a spot of an alarm clock. Once it shows up, I hold until I tin interact with Claude again, past archer it to grounds its advancement and extremity the session.

If I person much to do, I commencement a caller league and person it travel up to speed. This works well, due to the fact that the compaction process indicates that determination volition beryllium cognitive issues moving guardant unless Claude is allowed a refresh.

5. Use GitHub for security

Even if you're a lone coder, arsenic I am, you volition yet privation to link your task up to GitHub. In my Claude setup step-by-step article, I showed you however to acceptable up that link. Once the nexus is successful place, Claude tin perpetrate your enactment to the GitHub repo.

I similar to make a caller subdivision earlier opening enactment connected a caller capability, enactment done that capability, and past merge that subdivision backmost into the main repo erstwhile I cognize the capableness works. This is thing caller to programming. It's a champion signifier for immoderate benignant of programming, not conscionable AI-assisted work.

Also: Bad vibes: How an AI cause coded its mode to disaster

I really deliberation it becomes much important erstwhile moving with an AI, due to the fact that you ne'er cognize if the AI is going to bash something astonishingly destructive conscionable because.

To that end, there's a corollary, which is marque definite you backmost up your improvement machine. Since I'm besides investigating connected my iPhone, I marque definite I backmost that up arsenic well. Once again, you ne'er cognize if the AI is going to innocently bash thing enormously destructive, truthful having backups is good.

6. Build import/export capabilities aboriginal

If your task is going to usage a batch of information similar excavation does, physique an export and import capableness into the exertion early.

Coding information import and export is simply a reasonably straightforward process, but it tin beryllium time-consuming and tedious. As a result, galore developers don't codification it into their apps unless the quality of the exertion perfectly requires it. But with AI coding, it's nary much enactment than adding immoderate different capability.

In my case, erstwhile I started adding features to the app, particularly searching and filtering features, I needed the information successful spot to spot however it works. For me, those filtering features are cardinal to the program. If, for example, I privation to spot each the bluish filaments, I privation them sorted and presented to maine quickly.

Also: Coding with AI? My apical 5 tips for vetting its output - and staying retired of trouble

As it turns out, we (me and Claude) had to bash a batch of work, with a batch of revisions, to get speedy filtering. Having trial information was critically necessary.

When the codification goes wrong, arsenic it inevitably will, you mightiness person to delete the app and reinstall it. You don't privation to suffer each the information introduction you've done. So adding import and export capabilities is simply a large win.

Here's a extremity though: support successful caput that your interior information structures volition change. It's a bully thought to besides person the AI physique successful a information migration instrumentality truthful that each clip you person a caller information structure, it updates the aged data. Have it acceptable up a information operation version, and support updating the migration codification on with the import and export code. Trust me, you'll convey maine later.

7. How to deliberation astir debugging

The AI volition marque immoderate boneheaded moves. It's guaranteed. But it besides volition constitute tenable codification that conscionable doesn't enactment rather right, particularly successful the Apple ecosystem. Making the Quick Move feature, which required aggregate NFC scans tethered unneurotic successful a sequence, was a pugnacious project, requiring a ton of debugging.

After a while, the AI began to suffer its mode and I had to usher it. One of the cardinal things I had it bash was adhd interior logging, on with a mode to transcript that log from the iPhone into the terminal model connected the Mac for Claude Code. I besides had to archer Claude what to log, what to look for, and explicate what I was trying to accomplish.

Also: 9 programming tasks you shouldn't manus disconnected to AI - and why

It does adhd worth to the AI enactment to explicate precisely what you privation to do. For example, erstwhile I explained I wanted Quick Move for accelerated workflow, with a speedy pat connected a spool and a destination, it helped. But we kept moving into a time-out issue.

It wasn't until I explained that I mightiness bash a pat connected a spool, past person to locomotion to a caller destination and unfastened up a instrumentality (all of which takes time), that it realized it couldn't conscionable permission the default time-out settings. Instead, it had to bypass time-out settings entirely.

So, present are 3 main guidelines to support successful caput erstwhile debugging:

1. The AI is rather bully astatine debugging.
2. The AI needs guidance connected however to attack debugging, and however to stitchery the accusation it needs to diagnose a problem.
3. The AI benefits from knowing the existent usage lawsuit for which it is doing the debugging.

In my headcanon, I deliberation of the AI arsenic a precise astute grad pupil who tin bash the occupation well, but needs changeless guidance and wide instructions. That seems to enactment rather well, due to the fact that I support an oculus retired for points wherever it's wide the AI conscionable isn't getting the memo and needs to beryllium pointed backmost successful the close direction.

A corollary to this acquisition is simple. You request to test. A lot. Constantly. You, not the AI.

8. Organize your record operation

Claude tin beryllium a slob. It volition make files each implicit the directory structure, with nary respect for organization. Since it tin work and parse everything, it doesn't truly attraction wherever the files are, arsenic agelong arsenic they enactment to physique your app.

On the different hand, you and I are puny humans who request a logical operation to marque consciousness of anything. Additionally, adjacent if it's not astir making sense, conscionable having a agelong watercourse of random files strewn astir the task directory tin marque your psyche ache. It did mine.

Also: I asked AI to modify mission-critical code, and what happened adjacent haunts me

So aboriginal on, and past regularly passim the project, I had Claude spell done and restructure the directories, determination files, and reorganize things. Keep successful caput that each of these moves whitethorn necessitate a refactor, modifying the codification to recognize the caller locations.

So, don't conscionable archer Claude to cleanable up his room. You'll person to archer Claude what directory operation you want, oregon interact with the AI to get suggestions. Then aft each change, spell done and trial the resulting code.

Another tip: you tin archer Claude to analyse each the codification and fto you cognize if immoderate doesn't acceptable the existent structure. It volition bash that, and it tin drawback things, but it's neither 100% close nor 100% complete. So usage it arsenic a tool, but besides verify each change.

9. Claude is not the brag of you

Every truthful often, Claude Code volition get it successful its caput that it's the boss. It will, alternatively of doing the work, archer you however to bash immoderate needs to beryllium done. That mightiness beryllium record activities, oregon changing settings, oregon adjacent penning routines arsenic portion of the project.

The happening volition really respond to prompts by giving you instructions connected however it wants you to transportation retired the work. Do not springiness successful to the robot.

Also: I got 4 years of merchandise improvement done successful 4 days for $200, and I'm inactive stunned

Correct Claude, archer it that you privation it to bash the work. It whitethorn reason (make definite you person a backup), but if you insist, it volition really bash what you're paying it to do.

Sigh. Robots, these days. Right?

10. Use Swift and SwiftUI

Swift is Apple's open-source connection for gathering applications for the Apple ecosystem. SwiftUI is the related instrumentality for gathering the UI components connected those apps. While determination are immoderate different tools you tin usage to physique iOS and different Apple ecosystem apps, I powerfully urge you commencement with axenic Swift and SwiftUI.

Also: Claude Code makes it casual to trigger a codification cheque present with this elemental command

The crushed is that this attack allows you to widen your iPhone app to different platforms. For example, you could determination the iPhone app to the Mac, not arsenic an emulated oregon simulated app, but arsenic a autochthonal Mac app that shares immoderate of the logic, but separates retired the UI to beryllium much Mac-centric.

Because I built my app successful Swift, it whitethorn good beryllium applicable to usage overmuch of the existing codebase if I yet privation to marque a autochthonal Mac app oregon an Apple Watch app.

What astir you?

Have you tried vibe coding an app oregon important diagnostic with an AI similar Claude Code oregon ChatGPT Codex? What parts of the process worked good for you, and wherever did things interruption down?

Also: I'm an AI tools expert, and these are the 4 I wage for present (plus 2 I'm eyeing)

Do you hold with the thought of moving successful precise tiny steps alternatively than feeding an AI a afloat spec? How bash you grip documentation, mentation control, and debugging erstwhile an AI is doing astir of the typing?

And if you've built for Apple platforms, bash you deliberation Swift and SwiftUI are the close instauration for AI-assisted development, oregon person you had amended results with different approaches? Let america cognize successful the comments below.

You tin travel my day-to-day task updates connected societal media. Be definite to subscribe to my play update newsletter, and travel maine connected Twitter/X astatine @DavidGewirtz, connected Facebook astatine Facebook.com/DavidGewirtz, connected Instagram astatine Instagram.com/DavidGewirtz, connected Bluesky astatine @DavidGewirtz.com, and connected YouTube astatine YouTube.com/DavidGewirtzTV.