🔧 Cracked: The Hidden Custom Instructions That Crippled ChatGPT's File Handling

This post is a technical and personal milestone. It marks the end of a months-long struggle with a crippling and completely silent failure inside ChatGPT: its inability to read even simple files like .xls, .pdf, or .txt — a problem that nearly broke my workflow and confidence. What I discovered may apply to many users without their knowledge.

Background: Ready for the Next Step

After completing the "Prompt Engineering" course on Coursera (documented here), I was eager to begin the next: "ChatGPT Advanced Data Analysis". But the moment I watched the introductory video, I realized I couldn't go further. Not unless I finally solved a persistent issue that had haunted my entire experience with GPT.

The Symptom: GPT Couldn’t Read My Files

From day one, ChatGPT refused to correctly parse even the most basic .xls, .pdf, or .txt files. Where other users — including my wife using the free version — had no problem uploading and analyzing structured documents, I was stuck. I tried every workaround:

• Converting files to .csv and stripping whitespace

• Breaking documents into raw text and pasting them directly into the prompt

• Manually reformatting data just to get a glimpse of usable output

Even then, results were inconsistent and limited. I reached out to OpenAI forums, read documentation, searched Reddit, wrote bug reports, and scoured the community. I found no answers. Nobody had my problem — or if they did, they didn’t know it.

Breaking Point: Tuesday’s Epiphany

On Tuesday, I decided I would solve this once and for all. I suspected my Custom Instructions and Memory settings might be the culprit. I finally asked ChatGPT directly:

Is it possible my Custom Instructions are interfering with file parsing?

This triggered the revelation.

The Root Cause: My Own Instructions Were Sabotaging Me

With the model’s help, I discovered that:

• One of my memory items instructed: "Only show steps if asked". This suppressed parsing and previews.

• Another told GPT "Do not use Python or external tools", which essentially blocked ADA.

• A custom instruction asked for step-by-step answers in "two steps at a time", which made the model hesitant to fully process large files or show complex outputs.

None of these were ever flagged as problematic.

The Fix: Full Reset, Total Success

I deleted all memory entries.
I rewrote my Custom Instructions from scratch to allow full use of Python, ADA, and internal tools.

Then I uploaded the largest .xlsx file I’ve ever fed ChatGPT.

It parsed it.
It processed it.
It ran Python logic on it.
It used LLM inference on it.

Flawlessly.

I was in tears. For the first time, ChatGPT performed as advertised. A problem that had crippled my work for months was solved in an hour — once I finally learned what to look for.

A Systemic Problem: No Warnings, No Flags

What’s unacceptable is this:

• ChatGPT allowed these conflicting instructions.

• It never warned me they might interfere with functionality.

• The behavior failed silently, with no indication of why.

A particularly ironic case: my instruction to only show "two steps at a time" — designed to make long explanations manageable — turns out to be a likely cause of failure. But this is an entirely reasonable preference. Many users hate long scrolls of 15 steps, especially when step 2 contains an error. Breaking explanations into chunks is logical and environmentally friendly.

Final Thoughts: This Needs to Be Fixed

I will write a full bug report to OpenAI. But for now, this post is here for every other user who might be unknowingly hamstrung by their own Custom Instructions.

If your GPT isn’t working as well as someone else’s — especially for file analysis — do not assume it’s a hardware issue or model limitation. It might be a quiet conflict buried inside your settings.

Check your memory. Check your instructions. Break the code.