HELP! Any Tips on Prompts to Make the New 4o and 4 LESS AWFUL?

ChatGPT
, , , ,
1

Has anyone found effective ways for structuring prompts or giving specific instructions to ChatGPT to get better responses after the most recent rollout made versions 4 and 4o simply awful? I’ve been using GhatGPT’s paid version since it first became available. Since about April, I noticed drastic difference in quality of outputs. And, they have been getting worse.

I am spending SO MUCH time recently just tweaking and modifying my instructions to get even semi-intelligent responses. It is frustrating and feels like a waste of time. I compared the new outputs with my saved outputs from the previous version 4, and the difference is night and day.

Any tips or insights would be greatly appreciated!

11 Likes
2

I haven’t figured this out yet myself, but I feel your pain. My custom GPTs are acting completely randomly and giving awful output compared to a couple of weeks ago.

I found your post when Googling for info on new prompting structures.

Hopefully somebody replies with some insights.

10 Likes
3

Yes I get that too it has even failed to follow simple prompts in long threads and just goes all over the place

7 Likes
4

GPT-4o is different. This will happen again. Don’t see you tweaking as a waste of time. Learning to work through these bumps makes you a more skilled user. These skills are the gold of the future!

What is it doing that you don’t like? Have you tried to build a custom GPT? It’s pretty easy and would like you set it up to reply how you normal want it.

1 Like
5

I tell GPT-4o to cut all of its answers by at least half. Alternatively, you can set a token limit for its responses. Additionally, instruct it not to repeat anything it has already said. Include in the custom instructions to keep answers short, concise, and to the point, and add this to the “remember me” feature.

However, I’ve noticed that GPT-4o rarely follows the instructions in the “remember this about me” feature, so good luck with that. For me, it helps to explicitly tell it to cut all answers in half. When asking yes or no questions, I instruct it to answer with a “yes” or “no,” followed by a very short explanation. This approach seems to be working for me.

5 Likes
6

I didn’t find ChatGPT 4O that effective. However, after the launch of 4), even Chatgpt 4 starting throwing tantrum. I am thinking of saying bye-bye.

6 Likes
7

One tip is to stick with a model that you know works, ie a specific version of the model. Like for me gpt-4-0125-preview has been solid. When you don’t select a ‘specific’ (gpt-4o or gpt-4 for example) you will automatically get a new version without even knowing it.

In the mean time a lot of things keep changing, including in the API itself and then we tweak our prompts as well - so yeah its pretty easy to loose track of changes that might affect results.

(I WISH there was versioning in the Assistants, I change the prompt- its updated - no going back to what it was … - and to that point I wish I could also see what model a thread was run was in the back end)

I don’t know how you keep track of your prompt versions - but ideally you would keep the current production prompt, with the model version and every update you make. I have set of validation questions that I can run when considering switching models (or when I suspect openAI problems) - but I have more than 25 different assistants running and don’t have an automated system at hand for all of them.

3 Likes
10

Great advice. People should heed this advice.

To piggyback I also wish there we a version control because I’m doing empirical research so the stability of the model and making sure its the same model matters a lot.

1 Like
11

Provide worked examples in your prompt.
2-3 is usually plenty.

12

This is also a helpful tool to consider for keep track of responses across versions of models

2 Likes
13

Hi

I am very happy with the content I get from GPT 4o but it takes a lot to get it to write really good text that requires little to no editing.

If you’re getting poor results, get specific down to what each paragraph is. Get it to generate facts and cut out the fluffy content.

Here are some of the things I’m using now:

The role

“role” => “system”,
“content” => "You are copywriter specializing in TOPIC HERE. Give full answers that are packed with detail but do not have ‘fluff’ content.

A list of common AI words I want it to avoid:

$bannedWords = “adventurous and relaxing’, Nestled, azure, blend, charm, charming, cuisine, culinary delights, delightful, delve, dramatic, haven, history buffs, indulge, plethora, relish, renowned , rugged, seeking, transition, transitions, treasure, unique blend, vibrant, abound, abundant, allure, beckons, charming, comprehensive, convenience, cultural, dive into, diverse, ensuring, furry friends, gem, iconic, immerse, numerous, numerous benefits, option, recreational opportunities, rich, treasure trove, vibrant”;

The prompt:

You are an expert in [BLOG TITLE]
Write an in-depth, informative introduction to [BLOG TITLE] following the outline below.
Be conversational, use humour and relate to our audience which is [INSERT TARGET AUDIENCE]

Paragraph 1: About the location

  • An exciting intro like the one in the example text provided.
  • What’s this location best known for. Try to name 3 things.
  • How many Brits visit each year? Be specific. Include stats.
  • What’s new here? Be specific.
  • Explain why this is a good destination for families.

Paragraph 2: About this location

  • List the benefits of staying here? Try to name at least 3.
  • Why this location? Reliable brand, include the Feefo rating, 4.5 for service and 4.1 for product.
  • List pricing range for this location for a family pricing per week.

Paragraph 3: What to do on a holiday here

  • What are the best known, major tourist attractions here? Be specific and name them.
  • Name the best known festivals and the month they are on. Try to name 3.
  • Turn any attraction or festival into [attraction name] links

Paragraph 4: Practical details

  • How to get to this location. Some advice for anyone driving, very specific to this location. As well as some advice for flying. Inclide the airport and code.

Paragraph 5: Hook for reading this article

  • Make it funny, snappy, catchy.
  • We’ve ranked by user reviews

Guidelines

  • 150 - 170 words max length
  • Cut out any fluff or fluffy content.
  • Use

    tags for the paragraphs

  • Include humour.
  • Mention the location a max of two times, once in the first paragraph.
  • Be specific and use entities and real names of places, events. Always include real facts.
  • Include specific [location] activities available throughout the year by name.
  • Include real facts and entities.
  • Mention specifics such as the local towns, local food and the most popular nearby tourist destination.
  • Focuses on families thinking about booking
  • invite readers to imagine themselves in the location.
  • Strictly UK spelling.
  • Avoid all common AI content patterns.
  • Make the content sound human.

Strictly Don’t use terms:

$bannedWords

7 Likes
14

And since yesterday, I noticed it’s been freezing. Ugh…So tired of it. Just sent an email to their support. I doubt I’ll get a response but I still wanted to provide them my feedback and convey utter disappointment.

2 Likes
15

I have Prompt Engineering certification. The point I’m not trying to show off, but emphasizing that I’m pretty skilled at inputting decent prompts. I should not be modifying my prompts all the time based on terrible outputs. It IS a waste of time to me. It’s not making my work more efficient as before.

2 Likes
16

Good ideas! Thank you. I’ll try them. But, yeah, it’s funny you say “good luck with that” since I’m also not so optimistic it’ll make things much better, perhaps slightly but the quality of outputs themselves it may not fix.

17

I do that; still not making outputs better quality

18

Oh, I’ve done that. And, this is something I mentioned to OpenAI’s Support team this morning. I have screenshots of answers I received before the recent release. When I inputted the same question into 4 and 4o for comparison, the results were much WORSE than before.

1 Like
19

I’m glad it’s been working for you! I like that you provided it very detailed instructions. However, I noticed that even when I provide detailed instructions it still doesn’t follow them all the time.

1 Like
21

image

Has anyone had luck talking to a real person there?

22

Glad I am not alone experiencing this drastic awfully behavior in responses to prompts. However I always separate the actual with : and tupe my prompt in Capital letters, which just improves the tesponse. I hope they can get us back to the original responses type of model which was more accurate

2 Likes
23

Heh, the one time I got a human being in that support chat they promptly disconnected after 30 seconds and left me on unread for hours until I gave up

2 Likes