How to upload large code files to ChatGPT

Code is the easiest content to get into ChatGPT and the easiest to get wrong. Easy because source files are plain text and go through without fuss. Wrong because the instinct is to send the whole repository, and that reliably produces a worse review than sending a tenth of it.

Why the whole repository is the wrong move

There are two limits, and only one of them is technical.

The technical one is generous: 2 million tokens per text file, which is a great deal of code.

The practical one bites much earlier. Attention is finite. Give ChatGPT a hundred thousand lines and ask why a function returns undefined, and you have buried the answer in ninety-nine thousand lines of irrelevant context. Give it the function, its callers, and the type it returns, and you tend to get a real answer.

Flattening files into one upload

When the problem genuinely spans several files, put them in one text file with headers so the structure survives.

1. Pick the files that actually matter

   The one with the bug, whatever calls it, and any type or schema it depends on. Usually three to six files.

2. Concatenate them with path headers

   On Mac or Linux, something like:

   for f in src/auth/*.ts; do echo "=== $f ==="; cat "$f"; done > bundle.txt

   The header lines matter. Without them ChatGPT cannot tell where one file ends and the next begins, and it will hallucinate imports.

3. Upload the single text file

   One upload instead of six, and it counts as one against your rate cap.

That last point is worth noting: uploads are capped at 80 files every 3 hours, and on the Free plan 3 file uploads per day. Bundling six files into one is six times cheaper against that budget.

What to exclude, always

Everything here makes answers worse and costs you length:

• node_modules, vendor, target, and any dependency directory
• Lock files, unless the question is genuinely about dependency resolution
• Build output, minified bundles, and source maps
• Test fixtures and sample data, unless the bug is in them
• Anything generated

When the bundle is still too long

If your flattened file is past what fits, split it and send the parts in order, telling ChatGPT first that more are coming and not to answer until the last one lands. Otherwise it reviews the first file and stops. That is one of several ways to bypass the ChatGPT file upload limit, and the one that suits code best, since a bundle has natural boundaries to cut on.

There is a better option first, though: split it by concern rather than by size. Send the auth module and ask about auth. Then the data layer and ask about the data layer. Two focused conversations beat one enormous one, almost every time.

Asking better questions about code

The upload is the easy half. What separates a useful review from a generic one:

Say what you expected and what happened. "This returns null for users created before 2024, should return their profile" gets you somewhere. "Review this code" gets you a list of style opinions.

Include the error, in full. Stack trace included. It usually names the file and line.

Say what you already ruled out. Otherwise the first three suggestions are the things you tried yesterday.

Name the versions. Language version, framework version, runtime. Advice for React 17 and React 19 differs, and it cannot tell from the code alone.

A quick word on privacy

Anything uploaded goes to OpenAI's servers. For consumer plans, content may be used to improve model performance unless you opt out in the data controls. Business and Enterprise plans are explicitly excluded from that.

If you are working on anything proprietary, check your setting before uploading, not afterwards.