Why does my custom bot not follow instructions given to it?

I’ve been making some GPT custom bots where I give them a limited number of reference texts and instructions which is certainly within its 128,000 token limit so it should able to take everything into account when actually using it. Why does it ignore certain instructions given to it and how can I make it listen to them for every conversation? Please help, I would appreciate any advice!

3 Likes

Yeah instructions in my experience as well are unpredictable, not as stable as prompting. Also do make sure you’re not confusing ‘description’ with ‘instructions’, as that can be confusing.

3 Likes

Thanks for the response. Prompting seems to me much more effective too. Did you work out a system to make the custom GPT more accurate at using the instructions? I read somewhere that classification and numbering the important instructions helps but I am yet to try

This is why I am very disappointed with this race to have the largest context.
The models don’t actually “know” what’s in the context, and they only have a limited amount of attention heads that can pay attention to the context.
We already had the problem that the model couldn’t follow 50 rules at once back in the 8k context days; a 120k context with the same amount of attention will just be even worse at following the rules.

2 Likes

These responses are all very useful and appreciate all of them. I don’t have any programming experience so believed that I must’ve been doing something wrong. It sounds like it’s a common feeling around the forum. What would you all suggest for clarifying the reference text/context to the custom bot? Are there any particular effective tactics people have discovered? Are the instructions just meant to be a prompt that the bot continuously takes into account?

FWIW I’ve had a lot of success in just replicating the default system prompts (for GPT-4, browser, etc). Someone has extracted and hosted them here: https://github.com/spdustin/ChatGPT-AutoExpert/tree/main/_system-prompts

1 Like

So it’s a really good question @AliR - I have been studying this forum to learn as much as I can on the topic. So far what I have seen people say is…

  1. Format as Markdown - Markdown Cheat Sheet | Markdown Guide)
  2. Be detailed and specific, but not verbose. Clear succinct instructions help a lot.
  3. Include headers for each major function you would like the GPT to perform (how it should respond to common questions it will receive)
  4. Add very specific Conversation Starters (in the Configure tab) and provide details on how it should respond to each Conversation Starter
  5. Most well-crafted Instructions are between 1000 - 2000 characters. They can be longer if they are well organized and succinct.

There’s a lot more to add here. Maybe we can spin up a Best Practices thread as a quick reference for people.

3 Likes

Підскажіть. Завантажую книгу в файли свого бота. Завантадив та запитую чи бачить він цю книгу. Відповідь - ні; я не можу бачити вміст файлів. Чи можливо якось по іншому запитувати? Хочу щоб він міг зробити аналіз цієї книги та давати відповіді по ній, Чи файли для іншого? Дякую

"Перейдіть до редактора GPT за адресою https://chat.openai.com/gpts/editor/g-{gpt_id}, натисніть вкладку Налаштування та переконайтеся, що ви бачите файл, який ви завантажили. Якщо він там є, перевірте Інструкції. Додайте інструкцію, яка чітко говорить щось на кшталт "якщо користувач запитує про {файл}, звертайтеся до Знань у файлі {ім’я_файлу}.

Іноді просто потрібно змусити GPT зосередитися на Знаннях."

By the way, I had ChatGPT translate your post and my response from my native English so I hope the answer translates well.

I feel like there will be some problems lately. MyGPT used to end every message with a catword (Meow). Now he won’t speak at all.