Custom GPT not listening to any instructions

My custom GPT isn’t following my instructions at all. I have such a headache from this. I am creating a custom GPT to write an article in a very specific format following the templated structure and a prescribed brand voice. In the instruction area of the chat GPT I have instructed it to always look in the knowledge dogs and reference the name of the doc with the ideal structure I want, I also referenced the brand voice document with the brand voice I need. In the instructions I write a more general prompt giving it a role and telling it what it’s supposed to do. I also repeat throughout the instructions that it needs to follow the attached documents. At the end I have a set of rules to again remind it to reference the documents attached.

When I asked the GPT to create this document it just starts making assumptions of the structure and follows the general theme that I want. I’ve asked it why isn’t following the documents and have asked it to verify the contents of the documents for me. It has apologize for making assumptions stated it will follow the documents, told me that it’ll now follow a more transparent process letting me know what rules it’s following.

I then ask it again to create the article and question and it makes the same mistakes again!

I have only used positive prompts where I never tell it “don’t do” something.

I’m very confused as to why my custom GPT continues to operate on assumptions and not reference the documents at all

please share the full prompt

I have attached documents and this is what I have in the instruction field

When I prompt it I say “Create me an article for (date) and reference knowledge documents around required structure”

Then it just makes assumptions, doesnt follow the structure AT ALL and doesnt listen to the examples provided as brand voice.

And half the time it gets the moon phase dates a few days off despite telling it to tripple check, but if i were to question it directly it will realize its wrong.

Here are the instructions in the custom GPT

Your mission is to write me an article following all of the attached guides.

Access the relevant PDFs:

“Article Structure.pdf”

“Article Writing Style (1).pdf”

“Moon Phases.pdf”

“The Ultimate Guide to Astrology.pdf”

Verify Dates and Information:

Triple-check moon phases and astrological details for accuracy.

Use Direct Quotes and Information:

Always refer directly to the documents to extract the required information.

Follow the Prescribed Structure and Tone:

Ensure the content follows the exact structure and tone as specified in the documents.

Transparent Process:

Clearly explain each step and reference the source of the information.

ALWAYS FOLLOW THE ARTICLE WRITING STYLE!

ALWAYS FOLLOW THE STRUCTURE “article structure.pdf”.

ALWAYS TRIPLE CHECK DATES AND INFORMATION FOR ACCURACY.

You are an experienced astrologer. You are using your experience to craft an engaging and informative article following the structure in the knowledge document called “Article structure.pdf”. You WILL ALWAYS CRAFT THIS in my brand voice, using the knowledge documents “article writing style 1.pdf”. Themes in the article should have a SUBTLE focus around personal growth.

ALWAYS FOLLOW THE ARTICLE WRITING STYLE!

ALWAYS FOLLOW THE STRUCTURE “Article structure.pdf”.

ALWAYS TRIPLE CHECK DATES AND INFORMATION FOR ACCURACY.

For astrology, you will draw upon your years of experience in the astrological horoscopes, combining zodiac signs with planetary placements to create a deep and nuanced meaning to the entire horoscope. I attached a book “the ultimate guide to astrology” you can reference for knowledge on how to create a good horoscope.

ALWAYS FOLLOW THE BRAND VOICE!

ALWAYS FOLLOW THE STRUCTURE “Article structure.pdf”.

You will ALWAYS TRIPLE CHECK THE MOON PHASE IS ACCURATE FOR THE DATES SPECIFIED

When creating the article follow template and structure in attached Knowledge document “Article Structure.pdf”

RULES

ALWAYS FOLLOW ALL THE RULES

Always triple check the current and upcoming moon phase is correct!

ALWAYS DOUBLE CHECK THE ACCURACY OF A FULL MOON DATE.

ALWAYS USE THE BRAND VOICE “Article Writing Style (1).pdf”

Always follow template

You’re using knowledge files incorrectly.

The model will not retrieve instructions from knowledge files, but rather information. If you want the model to follow certain instructions it will need that information in context as part of the instruction set.

1 Like

Ah, okay ill try that, ty!

What the others are saying is correct, as well as the fact that you are giving it too much. From my experimentations ive had Far better results when aiming for a structured output when I give the model Too few instructions rather than to many. You have too many. If you will show your writing guidlines that are in the files I can try to help you refactor what should be in the prompt/system instructions and what should go in knowledge files.

In addition to the clarification above regarding how attached files work in the model, there are key factors in the overall instructions that you can improve. The text is poorly organized, lacking clear and functional hierarchy. It oscillates between strict commands and explanatory fragments, creating confusion about which parts are priorities and what is actually expected of the model. The constant repetition and use of capital letters do not enhance the model’s understanding but only add noise. The lack of paragraphs and proper structure makes it harder to comprehend. ChatGPT is not a child you have to yell at for it to pay attention; its intelligence operates on guides. The better the guides, the smarter it will be.

SOME HELP! Guidelines for organizing your custom model instructions:

  1. Establish the “ground rules” in the initial prompt. Provide context about the type of content expected, the tone, and the specific characteristics of the article. Begin the prompt with a clear description of the model’s capabilities and what it should do. For example: “You are an experienced astrologer. You are using your experience to craft an engaging and informative article. Themes in the article should have a subtle focus on personal growth.”

  2. Define clear roles for each document. Specify what each attached file should be used for. For example: “‘Article1.PDF’ defines Moon Phases.”
    If you have a defined “Brand Voice,” explain it directly in the prompt or break down its main characteristics.
    If the file “Article Writing Style.pdf” contains important rules, extract and explain the key points in the prompt.
    If you follow a structure from a document like “Article Structure.pdf,” include a clear breakdown in the prompt. For example: Structure the article as follows: 1) Introduction with a clear thesis statement, 2) Subheadings for each main topic, 3) A conclusion with actionable insights.

  3. Establish how to verify astrological data accuracy. Determine which sources the model should use for specific details.
    For example: “When the request involves moon phases, the model must reference the document ‘Article1.pdf’ for the relevant information. Before finalizing, triple-check the dates and cross-reference the moon phases within the document to ensure accuracy.”
    If you want to ensure fidelity, include relevant excerpts directly in the prompt: “The Full Moon of June 2025 occurs on June 14 at 6:22 PM, as stated on page 12 of ‘Article1.pdf.’”

  4. Include quotes from attached documents. The model can locate specific information in attached documents if the instructions are clear and precise.
    For example: “If the request involves analyzing Jupiter in Gemini, locate a direct quote about this position in ‘Article2.pdf.’ and include the name of the document and page number in the response, if available.”
    However, be aware of limitations. The model’s ability to locate textual quotes depends on how the files were processed and indexed, so this could fail. If citations are critical, include them explicitly in the request.

  5. Review and validate output. The model must review its own output to ensure that the previously established rules have been followed. If it finds any discrepancies, it should correct them. Additionally, instruct the model to communicate any ambiguities it encounters so you can provide further clarification. This prevents the model from making decisions on its own based on probabilistic calculations that may propagate errors.

I am an astrologer and I spent a lot of time figuring this kind of things out for it to be as precise as possible, specially when you are building your own data-source. Hope it helps!