Why AI coding gets worse as your app grows.

On day one your app was a few files. The AI could read all of them in a moment, so every request landed in the right place. Founders often read that early speed as a measure of the tool, when it was mostly a measure of how little there was to read.

Months later the app has dozens of screens, and the requests that used to take one message now take an evening. The slowdown tends to arrive gradually, which makes it hard to name. You may notice that a change shows up on one screen and is missing from another, that the AI builds something the app already has, or that you have started writing longer and longer messages to explain where things are.

None of this means you chose the wrong tool or that the app is beyond saving. In the apps we see, the cause is usually how the code is arranged, and arrangement can be changed without rewriting the product.

An AI coding tool commonly starts each new chat knowing nothing about your project. It finds out by opening files and reading them, and it can only hold so much at once. In a tidy project it opens the two files that matter and gets to work. In a disorganized project it is searching a junk drawer: it opens whatever looks close, guesses at the rest, and writes code based on the part it happened to see.

Duplicates are where this hurts most. Say the same Save button was written three separate times, once for each screen that needed it. You ask for the button to turn green. The AI finds one copy, changes it, and reports success. The other two copies never change. From your side the fix worked on one screen and failed on two, and that is why fixes seem to work about half the time.

The opposite arrangement has its own cost. When everything sits in one giant file, nothing is ever lost, and nothing is easy to find either. The AI has to read thousands of lines to locate the twenty that matter, and the longer the file, the more likely it edits a similar-looking part by mistake.

A kitchen works because the cutlery has a drawer and everyone knows which one. Code works the same way for an AI that has to find things from scratch every chat. Most small apps need only three labelled folders to get there.

The first folder holds screens, with one file per screen. If your app has a sign-in screen, a dashboard and a settings page, that folder has three files with those names. When you say 'change the dashboard', there is exactly one place to look.

The second folder holds shared pieces: anything used on more than one screen, such as a button, a page header or a pop-up. The rule for this folder is one copy, used everywhere. Change the button there and every screen that uses it changes with it, which removes the half-working fix described above.

The third folder holds the code that saves and loads data. Screens ask this folder for 'the customer's orders' and do not talk to the database themselves. When something saves wrongly, the fault is in a known place. The folder names matter less than the split; your tool may already have its own names for these, and it is fine to keep them.

Folders drift unless something keeps them in order. Since the AI starts each chat fresh, telling it your preferences once in conversation does not last. Most AI coding tools read a rules file kept inside the project at the start of every chat. CLAUDE.md and AGENTS.md are two common names for such a file, and your tool's help pages will say which name it looks for.

Put your standing rules in that file on day one if you can, and today if day one has passed. Four rules do most of the work: keep the app readable by a beginner, give each file one job, keep files short, and check whether something exists before writing it again. The block below is a starting set you can paste and adjust. Change the folder names to match your project.

After you save the file, open a new chat and ask the AI what rules it is working under. If it cannot repeat them back, the file is in the wrong place or has the wrong name, and that is worth settling before you rely on it.

You do not need to read code to know your way around your own app, and knowing your way around makes every later request shorter. Ask the AI to explain the folders in plain words: what each one is for, which files sit in it, and what would break if a given file were deleted. Ask it to avoid jargon and to define any term it cannot avoid.

Read the answer with the app open beside you. If a folder's description matches what you see on screen, good. If the AI describes two files that seem to do the same job, or cannot say what a file is for, write that down. Those are the places where duplicates and leftovers live, and they make a ready list for the next clean-up step.

Ask for the explanation to be saved into the rules file as a short map of the project. The next chat then starts with a description of where things are instead of working it out again, and you have a page you can hand to anyone who helps you later.