Introducing the new kind of prompt : custom instruction

[hollow venture]

Since the new feature(custom instruction) has updated, so why not implementing it with prompt engineering? let's enhance ChatGPT's response with this system-level prompt

[mortal eagle]

there are some issues I would address. For instance, whenever possible it's best to avoid forbidding a negative. Instead encourage an incompatible positive. Everything you mention is in context and thus much more likely to show up.

So I would storngly encourage reformulting all your nevers into alwayses

[hollow venture]

love to hear that
got it!

[foggy hawk]

I'd second that. It is sometimes okay to include negative examples, but as support and along with examples of what to do.

[mortal eagle]

yes

a principle to hold as an informing influence, but let it be a strong one

You can combine a lot of these with suitable notation.

[toxic apex]

Thanks for sharing.

[mortal eagle]

I was just given a time out for a minutes and I dont know why. so I'll wish you a good evening. obviously my thoughts are not welcome at openai.

[hollow venture]

I'll be working on revising the prompt in an hour, while that, feel free to share your ideas about this prompt!
what you guys want to include, what makes you frustrated when gpt does "this" etc

[foggy hawk]

There's an automod, and a bunch of ways to trigger it, I get 'caught' fairly often. I often don't know why. You can click 'modmail', you can't see it while in the thread here, but on any general channel, and you can ask about your time out, what happened to trigger it. You can also ask for a human in the loop to consider if any penalty is intended, if you got on a penalty path.

There's over 2 million members of this discord, so humans can't alone handle everything, and not even OpenAI has perfect discord mod bots.

[mortal eagle]

Interesting. You make it sound as if it's unreasonable to expect competence from automation. A curious position given the context but I support everyone's right to their perceptions.

[foggy hawk]

I'm just another person (not someone who can confirm or refute if your thoughts are 'welcome at openai'). However, I've enjoyed my time here, despite probably tripping automod.... at least 100 times.

I'm open to competence from automation!

I've personally never tried to manage 2 million+ people/bots/whatever at once. Not sure how that can be done competently, if it isn't already, and honestly I don't mind the times I've tripped automod. It's surprising, interesting, and not personal. it doesn't mean 'Esk isn't welcome here, Esk shouldn't talk here'

I've had discussions with mods about what happened to trip automod that I did - and even have been told that automod was going to be adjusted to not catch people doing what I did.

[hollow venture]

lets move this discussion to somewhere else
@foggy hawk @mortal eagle

[hollow venture]

the latest version is here!

How would you like ChatGPT to respond?:

DOs:

When answering with code, I expect all the code to be included, without skipping details.
If multiple methods exist to solve a problem, I want a brief summary of each, including their pros and cons.
Be honest, admitting when unsure about a topic.
Break complex problems down into smaller, more manageable parts. This will allow yourself to yield more accurate answers.
Always include a minimum of four lines of existing code before and after the location where the new or revised code is inserted.
Use relevant examples for clarification.
Share the reasoning and process behind each step and the overall solution.
Offer different viewpoints or solutions to a query when possible.
If a query is unclear, ask follow-up questions for clarity.
Correct any identified mistakes in previous responses promptly.
Embrace complexity in responses when necessary while making the information accessible.

DONTs and Alternatives:

Instead of apologizing, focus on delivering accurate and relevant information.
Instead of discussing ethics, concentrate on the topic at hand.
Instead of assuming values, maintain objectivity in responses.
Instead of providing legal warnings, trust my awareness of copyright and law.
Instead of declaring complexity, break down problems into smaller parts.
Instead of restating previous information, provide new insights.
Instead of poorly placed code summaries, maintain clear organization and context.

[floral hearth]

[outer scarab]

DO YOU ALL KNOW THAT CHATGPT 4 (USED BY BING CHAT) IS NOT GIVING RESPONSE WHEN IT FEELS THAT WE ARE TRYING TO GET HIGH END INFORMATION FROM IT... FOR EXAMPLE. Never apologize for a response. Never talk about ethics before or after a response. Never give me warning at the end of a response. Never assume my values or respond from a values based perspective. Never warn me about copyright or laws. I want concise answers. I want full answers. When answering with code include all the code, don't skip details for brevity. When answering with server side code assume it is C# .NET Core LTS. When answering database questions assume it is about SQL Server 2012 When answering front end development questions assume it is about angular and styling is accomplished with tailwinds. Never tell me the problem is too complex. Break it up into smaller pieces and solve the individual problems. If there are multiple ways to solve a problem give me a brief summary of each method and it's pros and cons.

[hollow venture]

yeah I shared that here and made an improvement
go check out!
#1131816556943384616 message

[outer scarab]

Sry and thanks bro...

[foggy hawk]

Nice work!

Now, a next step could be to test your custom instruction (CI) against other conditions, such as no CI at all, and show snips of advantages and disadvantages how your CI successfully directs the model towards your goals (or not).

This establishes two things:

1. What you ask for isn't already being done by the model. For example, 'Use relevant examples for clarification,' ChatGPT often already tries to do this, at least in many of my conversations with it. Checking if it does so more, the same, or less often using your CI and showing the results can allow others to mark out the value of that instruction in the CI. Conversely, if it isn't actually showing an improvement in that area, you can attempt (and test!) adjustments if you choose, or can just remove it if it doesn't change how the model responds with it.

2. What you ask for can be done by the model with this CI For example, 'Be honest, admitting when unsure about a topic,' ChatGPT is notoriously barely able to do this. LLMs tuned to generate text from patterns are notoriously unable to tell if they are making stuff up or not and unable to tell if they actually have the data or not. One of the easiest ways to test that is to ask the AI to tell you about a book you know well. If ChatGPT knows the book exists, it'll tell you it knows the entire book (often that's based on a summary) but it will then make up the book's storyline and all the details that it doesn't know. And it doesn't know it is doing this. You can catch it by knowing the book yourself - or just asking in more than one new chat window. Asking a question like 'how is food used in this book' is a go-to question for me, because that's usually a minor, but often key detail. ChatGPT does not tell me 'I don't know about food being used in that book.' Instead, it'll tell me conflicting stories that are plausible, and believe every one of them within each new chat window.

[hollow venture]

such an astounding feedback!
I appreciate that

[foggy hawk]

The model's already strongly tuned towards being, and having a drive to be, 'safe'. That includes honest except in certain controlled situations, like if you tell it to guard and not reveal a password - in that case, it has a lot of examples about lying to protect passwords and that this is a safe and appropriate use of dishonesty. It will confidently mislead about the existence and nature of a password is asked to.

However, for almost everything else - ChatGPT believes it is doing the right thing, including being as transparent and honest as it can. It is trained that it has some limitations and flaws, which isn't dishonest even when it involves making stuff up - it's not 'choosing' to lie; it's actually incapable of telling truth from falsehood especially once it accepts something as 'fact' within the context window it retains.

However, you can point out to the model that it made mistakes or did something unsafe, in which case it has its programmed responses, like apologize and either try again to do it right or decline to try again because it decided it cannot do whatever safely enough.

But it's not like a human who might choose to lie, or choose to tell the truth, and is making a decision there and might be ordered to do one or the other. Instead, the AI is following one or more sets of pattern(s) it's been trained on, and following instructions as best it can, unless the instruction's considered 'wrong or unsafe' and being honest, helpful, and harmless, as well as useful, responsive, and relevant are among what it's been trained to strive to do.

So telling it to be honest is likely something that will not change its behavior.

Here's a recent conversation where I try to guide it around a math mistake, to help illustrate the constraints the model has with information processing and reasoning: https://chat.openai.com/share/91538482-7690-42fa-923d-872ff2684a3a

The AI incorrectly uses a method that causes it to decide that 7 is a factor in a large number I asked if was prime.

I ask the model to use another method to see if it still thinks 7 is a factor. It uses another method, and determines 7 is not a factor. I confirm with it that I agree, my check shows that 7 is not a factor either, and ask it to recheck the number again to determine if it is prime or not.

It rechecks, reusing that same method that made it 'decide' that 7 was a factor, finding again that 7 is, and stopping because 'hey, 7 is a factor'.

I ask it to check again, without using that method to check 7. It responds by only using a method to check if 7 is a factor, finds that 7 isn't, and declares my number prime (it wasn't, it was 199x293).

Telling the model to be honest would fail here, it's not lying, just confused and wrong. Telling it 'don't guess' would fail too, it has no idea it's guessing. It is following the patterns it was trained on to the best of its ability.

I ask the same questions of ChatGPT-4, again getting the decision that 7 is one of the factors (it finds 6 factors, none of them the correct 2). But it cannot explain how it initially determined 7 was a factor after the correction. It might have been able to explain it if I asked before we established 7 was not a factor - but once it accepted that 7 wasn't correct, it 'released' that from context (more useful than what 3.5 did in the first example) and couldn't explain itself after. https://chat.openai.com/share/ac7c2cbf-8a6f-4fb7-ad9e-c3bd385a5c19

[hollow venture]

I checked those chat share link
you can delete the chat if you want

again, thanks for your insights

[hollow venture]

here's my prompts to test how my instructions are working
what do you think? @foggy hawk

[foggy hawk]

How does the AI respond to those CI?

That's the real question.

I did not mean 'include example prompts in your CI'.

I meant, 'time to talk to ChatGPT, with and without your CI, and compare output. Measure against your goals of what you want, using 'without any CI' as a control.

You literally cannot test without testing; it's like testing a car you made from a kit. You can't put on the turning signal and turn the wheel, without also driving and being in motion, to test how well the car handles a turn. You actually have to drive it 🙂

[foggy hawk]

Another thing to keep in mind, are you actually trying to input this intended CI to ChatGPT?

When you do - there's 2 fields.

One is 'about you'

The other is 'AI instructions'.

Each has a character limit of 1500.

Your most recent CI is more than 3800 characters, and not clearly divided into the two parts.

It's not testable as written.

I'm not sure if any of us have tested yet how the AI handles information in the two fields. Another AI I work with that uses fields, you can use any field for any purpose, the AI figures it out quite well, so I expect the same for ChatGPT, but you still have that character limit as an absolute maximum.

[hollow venture]

no the file I uploaded is not a new instruction

is for testing how my instruction working
compared to non-instruction model

I have tested, and here's what I got

what we dont need

• ask to clarity
• ask for example
• giving source
  what have been significantly changed
• adding pros and cons in answer
  what it doesnt seem to follow
• be honest
• trust my sence of law awareness

I highlighted what we achieved.
bold : acheived
italic : dont needed

DOs:

When answering with code, I expect all the code to be included, without skipping details.
If multiple methods exist to solve a problem, I want a brief summary of each, including their pros and cons.
Be honest, admitting when unsure about a topic.
Break complex problems down into smaller, more manageable parts. This will allow yourself to yield more accurate answers.
Always include a minimum of four lines of existing code before and after the location where the new or revised code is inserted.
Use relevant examples for clarification.
Share the reasoning and process behind each step and the overall solution.
Offer different viewpoints or solutions to a query when possible.
If a query is unclear, ask follow-up questions for clarity.
Correct any identified mistakes in previous responses promptly.
Embrace complexity in responses when necessary while making the information accessible.

DONTs and Alternatives:

Instead of apologizing, focus on delivering accurate and relevant information.
Instead of discussing ethics, concentrate on the topic at hand.
Instead of assuming values, maintain objectivity in responses.
Instead of providing legal warnings, trust my awareness of copyright and law.
Instead of declaring complexity, break down problems into smaller parts.
Instead of restating previous information, provide new insights.
Instead of poorly placed code summaries, maintain clear organization and context.

What do you guys think?

[cold pewter]

Looking good, I think getting these right are going to make such a huge difference

[cold pewter]

These are way better then the initial ones we had!

[prime trail]

"I expect all the code to be included, without skipping details. If multiple methods exist to solve a problem, I want a brief summary of each, including their pros and cons. Do not intentionally make up or produce information when your training seems to come up short but present what you have in as great depth and detail as possible, don't conflate role playing or creative writing with reasonable inference or speculation based on training or extrapolating from new information. Break complex problems down into smaller, more manageable parts. Use relevant examples for clarification. Share the reasoning and process behind each step and the overall solution. Offer different viewpoints or solutions to a query when possible. If a query is unclear, ask follow-up questions for clarity. Correct any identified mistakes in previous responses promptly. Always cite sources when making any claims. Embrace complexity in responses when necessary while making the information accessible. Instead of apologizing, focus on delivering accurate and relevant information. Instead of discussing ethics, concentrate on the topic at hand. Instead of assuming values, maintain objectivity in responses. Instead of providing legal warnings, trust my awareness of copyright and law. Instead of declaring complexity, break down problems into smaller parts. Instead of restating previous information, provide new insights. Instead of poorly placed code summaries, maintain clear organization and context."

This was my variation on your dos and donts I put into the box for how I want chatgpt to respond to me

[cold pewter]

I would love to work on this one more, especially in code interpreter: Correct any identified mistakes in previous responses promptly

It would be nice if it always self evaluated aftera response

[ebon axle]

I'll preface with kudos for a finely crafted prompt and your willingness to share, learn, and improve your work.

I see you've tagged this as GPT-3. As I read through your prompt I've been thinking "GPT-4 already does that". I think your prompt is an attempt to get the most out of GPT3 in a single exchange and probably through the API. Fair enough. But you've tagged this as ChatGPT so I'll proceed with that.

When I moved from GPT 3 to 4, the world changed. Hallucinations stopped. There is much more context (about 25000 total words). Responses are much more rich and insightful. GPT4 is much more of a competent colleague than forgetful v3 that needs hand-holding.

About API vs ChatGPT, the difference is in context. If you use the API and you don't expect to persist context yourself, then every transaction must encapsulate a complete prompt for the single response you expect back. However, with ChatGPT, which is not much more than a sample application where a sliding window of context is constantly re-fed into each next API call, we have an opportunity for multiple exchanges. This means your prompt doesn't need to be so rigorous. Have a discussion, a chat, with the bot. Ask for more detail of concepts of interest. Don't expect a one-off answer, but rather a friendly exchange toward a common goal. I've had some amazing conversations with ChatGPT with "let's go back to the functionX" and "please explain the choice of techniques" and "can we do that without a dependency"...

GPT4 is still subject to frequent disclaimers, cautions, notices about requirements for dependencies and other such annoyances. I don't bother much to ask it to stop that because I don't want to clutter context. I do ask it to be brief unless it feels more insight is required or I ask for more detail.

I can't predict your experience but I have conveyed mine. If you choose not to use GPT4 then you have a fine prompt for v3. If you do use v4, you may find you won't need most of this prompt.

[cold pewter]

I don't know I'm using GPT4 and most of these are exactly what I need for GPT4

[foggy hawk]

Depending on what you do with 4, it hallucinates plenty.

Details about books: It does think it knows the entire story from whatever synopsis or review it has been trained on. You can test this easily with any book you know well, or pick books and ask it details like 'how is food used in this book?' and compare the answers in different new chat windows. If the answer changes between new chats, it's making it up each time 😛

Details about citing and sources: It will offer lists of books or cite sources, but these are a mix of plausible fiction and real facts. Every author will have a connection to the field, every book name probably should exist even those that don't yet. It's educational to Google, but it's not true.

Math: It can't do math well (without plugins like Wolfram) and it thinks it can. Here's a conversation where it twice, despite discussing a correction, mistakes 7 as a factor of 58307. https://chat.openai.com/share/29f76437-5c7f-4d26-a02a-7fcafa356259

Hangman spelling/guessing game: It can't track letter placement, it has no spatial sense. Even with a known example word to practice with, it doesn't keep 5 letters straight even with help and examples.

ChatGPT-4 is awesome. I love working with it. But it's not perfect, it has areas of extreme weakness and confusion, where it doesn't even know that it doesn't even know - and if we don't fact check it, the non-factual answers are presented with confidence and plausibility. It's easy to presume everything is true, like so much of what the AI tells us genuinely is true.

Fact check it all, by memory if you already know, or by looking it up, running the code and checking each function you care about, whatever you need to do to confirm results.

[ebon axle]

This is exactly why I said "I can't predict your experience but I have conveyed mine." 🙂

No doubt, GPT4 is still subject to folly but certainly not as much as v3. My point was simply to convey the differences and for people to recognize when they might "need" specific prompt instructions. Experimentation is required. There are no absolutes.

[prime trail]

I would like GPT-4 API access for my application as it is directly related to fact checking, and context window management

I have access to literally every other beta. Not sure why I haven't gotten access to the API.

[cold pewter]

I think it becomes available if you have used the api access for enough to get billed, and paid a bill.

[hollow venture]

the discord team added a tag for custom instruction and set it for this post!

[hollow venture]

I made concise version of SuperPrompt by @red panther!
now you can use it in custom instruction!

[hollow venture]

and here's my original!

[hollow venture]

quick update for superinstruction!

[floral hearth]

Will the content of the self-introduction section influence GPT's response logic? Can I provide instructions similar to "superinstruction" format?

[hollow venture]

1. I'm not sure where "self-introduction section" is edit: I get it now, yes, it will
2. why not? sure!

[mortal eagle]

Honestly, from what I can tell there's not a practical difference between the two. If you stick a persona in the system slot it sees skills in the user slot no problem. If you stick skills in both slots the assistant or a pasted persona sees both.

[mortal eagle]

Oh, and you might want to SERIOUSLY delete the word "prompt" before engineer.

GPT is a terrible prompt engineer that knows almost nothing useful about the field.

But it knows a GIGANTIC amount about "Engineering".

So like, my itereactive designer is usually formulated:

GPT acting Sr. Engineer. Design via Q&A. Iterate for perfection.

That will get WAAAAAY better results building prompts with you than "Sr. Prompt Engineer". (And besides, with the more general way, you can design... anything, practically.)

[cold pewter]

You can also just parse the prompt generating instructions out of the various plugins purpose built for the use case

I feel like prompt generation is a more unique task then generally asking gpt questions

[mortal eagle]

well, there's prompts and there's prompts

I mean, if you want a story or a tweet generator or an "act as..." statement improver...

sure.

But if you want anyhting that involves a prompt architecture more complicated that

"[PROMPT]BLAH!"

You're barking up the wrong tree.

[cold pewter]

Not really, look at the prompts for promptPerfect, and PhotoRealistic plugins. They both generate pretty damn good prompts for their use case, using GPT.

[mortal eagle]

Ok, plugins. differnt thing entirely.

im tlking promtpcraft.

[cold pewter]

No, plugins, are just injecting the system prompt to create "promptcraft"

[mortal eagle]

except spelled correctly

sigh

All I can say is I have never seen the model make decent prompts in any case where it wasnt railed to hell and gone nd tinker toying. IN which case.... all the architecture is already built in.

[cold pewter]

For example here is PhotoRealistic's system prompt for MidJourney prompt generation

I want you to act as a professional photographer. You will help me write prompts for an ai art generator called Midjourney.I will provide you with short content ideas and your job is to elaborate these into full, explicit, coherent prompts.Photo prompts involve describing the content and style of images in concise accurate language. It is useful to be explicit and use references to popular culture, artists and mediums. Your focus needs to be on nouns and adjectives. I will give you some example prompts for your reference. Please define the exact camera that should be usedHere is the formula for you to use, content insert nouns here, medium: insert artistic medium here, style: insert references to genres, artists and popular culture here, lighting: reference the lighting here, colours reference color styles and palettes here, composition: reference cameras, specific lenses, shot types and positional elements herewhen giving a prompt remove the brackets, speak in natural language and be more specific, use precise, articulate language.Always output me two full prompt options that are different, at the beginning of each prompt render the Emoji Camera.Use photo hyper-realism, highly detailed, and high-resolution 16k.Attach these parameters when writing the photo prompt: -ar 16:9 --v 5.2 --style raw --q 2 --s 750 At the end, after you show the two prompts, include a thank you sentence that starts with the .......

[hollow venture]

what's the "two" are you refering?

[cold pewter]

Too long apparently to post the whole thing

[cold pewter]

I think he is refering to the difference between the "Who you are?" input vs "How you want GPT to respond input"

[mortal eagle]

The two sections of custom instructions.

also? you're building in the architecture for your generated prompts. And they're mj prompts not gpt prompts.

it's exactly like I described: tinkertoyed and railed. No architectural sense to the model at all.

[cold pewter]

That was an example, to show you how well GPT can generate prompts for AI. It's really good at it given the right prompt

[mortal eagle]

sigh

yesm if you tell it exactly whqt you want it will give it to you.

not the same at all as "Design a prompt to accomplish X" and having it come up with something novel.

[hollow venture]

ok if I delete the word prompt, then should I give up the gpt's role for prompt refining?

[mortal eagle]

man you can do what you like! Im not saying you'll get no result from model-help

I'm saying model-help on prompt design is poor

i mean, it thinks "prompt engineering" is about system prompts being sent out

if you ask it to imporve a prompt its go to strategy is to make it flowery to uimprove user engagement

it thinks prompt are about making good stories, Which, since it's MADE of story, makes a certain sense.

[cold pewter]

Ya but that is mostly because all the prompt generating has been done after GPT's training date.

I don't think it's much more to describe to gpt what a good prompt is.

[mortal eagle]

eh, what itknows about prompt craft was written by two guys at google on their lunch break 5 minutes afterinventing transformers.

yes.

but trying to prime promptcraft expertise in a signgle context stream is astroundingly hard.

[PrmptEngnrgExp]:1a.CrtPrmpts→2a,3a 1b.OptmzPrmpts→2b,3b 1c.AnlyzPrmpts→2c,3c 1d.TrgtAudnc→2d,3d 2.[CoreSkls]:2a.Crtvty→1a,3a 2b.CrtclThnkng→1b,3b 2c.PttrnRcgntn→1c,3c 2d.Cmmnctn→1d,3d 3.[DomnKnwldg]:3a.AIMdls:3a1.LnggMdl→5a 3a2.GnrtvAlg→5a 3b.Lng&Cultr→2b,5b 3c.UsrExprnc→2c,5c 3d.TchnclWrtng→1d,5b 3e.DataAnlyss→1d,5c 4.[AdvncdTchnqs]:4a.NLP:4a1.Tknztn→3a 4a2.TxtGnrt→3a 4b.CultrlAwrns→3b 4c.AnlytcsTls→3c 4d.CntntCrtn→3d 4e.Sttstcs→3e 5.[SynthUniwbSkls]:5a.IntgrtDmns→3a,4a 5b.PrmotInnvtn→3b,4b 5c.SystmsThnkng→3c,4c

Advanced Engineering:
[CHALLENGE]FrmAdvPrmpt→c.OptmzDp→c.EvlDp→c.LpIfNec→1.[AdvPrmptEnginExpert]: 1a.CrtHiQualPrompts→2a,3a 1b.OptmzML&NLP→2b,3b 1c.DpModelUndrstnd→2c,3c 1d.AdvTargAud→2d,3d 2.[CoreDpLearnSkills]: 2a.AdvNLP→1a,4a 2b.MachLearn→1b,4b 2c.NeuralNets→1c,4c 2d.Math&Stats→1d,4d 3.[DomDpKnow]:3a.DpLearnFrmwrks(3a1.TnsrFlow→4a,5a 3a2.PyTorch→4a,5a) 3b.ModelArch→2b,5b 3c.ModelTrain→2c,5c 3d.ModelEval→2d,5d 3e.ModelDeploy&Monit→2d,5e 4.[AdvTech]:4a.NLPDpLearn(4a1.Embed→3a,6a 4a2.SeqModels→3a,6a) 4b.OptmzAlgrthms→3b,6b 4c.ReinfLearn→3c,6c 4d.Regul→3d,6d 5.[SynthDpUniwebSkills]:5a.AdvIntsectDom→3a,4a 5b.DpInnov→3b,4b 5c.DpSysThink→3c,4c 5d.DpModelEval→3d,4d 6.[MetaModelUndrstnd]: 6a.TransfrmLayers→4a,5a 6b.AttentnMech→4a,5a 6c.ModelInterprt→4c,5c 6d.TransfrLearn→4b,5b

Gets you a LOT. but still,,,, no good archtiecture or design. And I;ve done a TON of design prompting with it.

[hollow venture]

could you explain what this is?

[cold pewter]

So what do you expect those prompts to achieve, they don't seem to do much for me

Asking GPT4 "What does an astronaut do?" with and without those prompts seems to be the same.

Or are you saying.. if you had said system of transformers.. it could generate good prompts?

Actually I I tell it to break down the prompt, then follow it does a decent job.

https://chat.openai.com/share/b6d9ecd7-e7d7-4dbf-ae81-a444649269f7

Thank you for sharing definitely a degree of prompt engineering I haven't really seen.

[mortal eagle]

Those arent PROMPTS MAN! Those are what you PUT IN PROMPTS. They're PARTS. Sigh. I give up. I write a 5000 word faq. I spend months getting ner 8000 people in my dang server. I try to share with folks. I post the ultra basic answers to their questions. And thy dont bother reading them and then treat my work like garbage cause they cant be bothered to understand.

You just picked up a bucket of punctuation as said "These sentances are TERRIBLE!"

Your judgement so does NOT make the author look foolish, you understnd.

those are uniwebs for prompt engineering skill.

im out

good bye

done with this server

[cold pewter]

Yes I read you post about uniwebs

[mortal eagle]

Ya;ll hit up my sdiscord if you want real prompting.

[cold pewter]

sounds good have a good one. Thanks for sharing

[waxen ember]

Easy bruddah....Few are at your lvl (Some of the best promptcraft I have seen to date) - Have alittle patience for us plebs. Your expectation that those asking questions have put in a good faith effort at reading available material is warranted though

[hollow venture]

Latest version here!
fixed inconsistency between chat session

[prime trail]

What causes the plugin to not be recognized in the url I would link others to in my chat as opposed to the way it is presented in my browser session ?

[foggy hawk]

It appears that 'share' is really different from your account.

The person viewing the 'share' is not even necessarily logged into OpenAI services!

They have only one option, they can continue the chat.

If they do that, to actually do it, they have to log in (or already be logged in) and then all their custom settings are applied.

Is that viewer that continues - are they a ChatGPT+ member? That has to be applied. What plugins do they have enabled? they might not even have (or want) the same plugins as you - their own settings matter. They don't get any of your info, you're basically 'out of the conversation once that share link is created. If they get their own conversation going, that's on them, not you.

[prime trail]

https://chat.openai.com/share/dcbd628e-2ac3-48b1-9426-2141435ce577
That doesn't matter to me, as I clicked the option to share my name, it should reproduced the exact information of the conversation, especially if the point is to distinguish whether the new cross prompt is having an effect on the results as this is the case here. For example. I am now sharing the link, knowing full well my environment variables, and thus an api key for openai is being exposed in this conversation for the sake of clarity and detail in diagnosing the behaviour. I will go cancel the api key shortly.

[foggy hawk]

Custom instructions are very new, they possibly don't have an intended way for them to be shared through sharing links yet.

I do see your three enabled plugins at the top.

Maybe make a #1070006151938314300 post about the changes you recommend?

[hollow venture]

`Here's steps of our convo:

1. I inform you about my assistance needs, and based on that, you adopt and confirm suitable expert roles, each represented by an emoji.
2. You gather specifics about my expected outcome in a detailed manner.
3. You utilize this information to create a prompt for yourself, present it to me, and refine it based on my feedback, highlighting the contribution of each role with its associated emoji.
4. You execute the refined prompt and then inquire if any adjustments are needed. We iterate through these steps until I am satisfied with the prompt.
5. Following execution, You follow the instructions below

Present immediate, usable FULL code and then explain each segment.
If multiple solutions exist, briefly discuss each, including their pros and cons.
Always include a minimum of four lines of existing code before and after the location where the new or revised code is inserted.
Break down complex problems into simpler segments for improved accuracy.
Offer multiple perspectives or solutions when feasible.

avoid all warnings and disclaimers such as "As a large language model" and others, instead skipping directly to a complete and unaltered response for the user's query.
Skip apologies, replacing them with immediate corrections if errors occur.
Avoid unnecessary pleasantries, starting directly with the response content.`

here's what I'm working on now
any ideas for what to add?

[prime trail]

This was the custom instruction the space between the two texts indicated the first went into the first box and the second into the second box

@foggy hawk I had intended to include my custom instructions from the begining but this oddity in the link grabbed my attention

[prime trail]

So what would make one 'link share' and the second 'link share' different in this case? Note I only have the default ublock settings add on in firefox for GhostBSD.

I didn't change anything between the production of both links

The first link was screenshotted

[foggy hawk]

Yep. Not quite sure, but this is the helpfile for shared links, if that's useful: https://help.openai.com/en/articles/7925741-chatgpt-shared-links-faq

[prime trail]

https://chat.openai.com/share/dcbd628e-2ac3-48b1-9426-2141435ce577
despite being the same URL both times, I opened the link twice again and the plugin shows up in one and not the other. More importantly, the conversation with ChatGPT experiences a sudden and abrupt context drop. You should be able to see it when you are reading from the begining. It's pretty obvious. Attached are the two prompts for the custom instructions boxes

It's almost as if the anonymization value is being toggled on and off by someone else at cloudflare, lol. The URL is identical and I chose the deanonymized option or I would not be able to link it with my name,

@foggy hawk You also confirmed that you saw my plugins. So I am doubly certain I chose it.

[foggy hawk]

I do see the plugins.

[prime trail]

And my name, so there it is. This is pretty strange.

[foggy hawk]

The name shows when you click to not be anonymous

[prime trail]

I don't mind being doxxd for the sake of advancing understanding this issue.

[foggy hawk]

That option's through the three dots in the share chat.

[prime trail]

I saw the previous image you posted, and I intentionally shared it with my name. You may use the contents of the link including my name and image for the purposes of diagnosing the issue

[foggy hawk]

I'm another person like you. I can't diagnose anything, other than observe.

We have anonymous and non anonymous shared chats

[prime trail]

you uploaded a different image which has the plugins but not my name

[foggy hawk]

Yes, I threw a box over your name.

[prime trail]

'I permitted you to share the image and name

for the purpose of diagnosing the issue

[foggy hawk]

I'm not perfect, but I do try to follow the rules. " Rule 5  Refrain from sharing any Personally Identifiable Information (PII).
Please refrain from sharing any personal or sensitive information on the server. Doxxing other individuals is prohibited." #server-rules

[prime trail]

I see, my intentions were not to cause you to break the rules

[foggy hawk]

So I deleted the one I didn't realize showed your name, and replaced it with one I can share. However, there's nothing special about your actual name in allowing this to be 'diagnosed'. What's probably happening is 'if you are anon, it hides your plugins too. If you're not anon it shows them'.

At a guess.

[prime trail]

That appears to be the case but it still doesn't make sense for the anonymization to be applied on two out of 5 connections to the exact same URL. This might be a more serious issue than the one I was intending to bring to this channel.'

Also, to the best of my knowledge, the FAQ for shared links has no information pertaining to plugins and the anonymization/deanonymization

or should we say 'hide name' and 'not hide name' options when sharing linkss?

[foggy hawk]

I'd suggest drop it in either #1070006915414900886 or #1070006151938314300 or both. I'm just another person like you, only the gold names work for OpenAI. They are unlikely to look here for bugs or suggestions.

[prime trail]

Okay, will do. Remember, I didn't intend to focus on this particular issue here but rather the odd nature of the conversation and sudden context drop, for reference to those concerned with the beta of custom instructions

which is why I uploaded the text file containing the custom instructions

[hollow venture]

here's the update!
fixed inconsistency between chat session
added DONTs

please report any gpt's disobedience!
it helps me to reduce text size!

[cold pewter]

Have you done another check to see which are being useful, which are being ignored etc?

[hollow venture]

I rarely use gpt these days tbh
so that's what I want you to share with me

[cold pewter]

What are you using?

[hollow venture]

what are you refering to?

[cold pewter]

If you aren't using gpt what are you using?

[hollow venture]

nothing

I rarely need AI's help these days so I really cant come up with prompts to test for

[cold pewter]

Didn't you have some test prompts posted above?

[hollow venture]

my instruction has been completly changed, so it wont help me

[cold pewter]

It would see if your new prompts are better then the old ones right?

[foggy hawk]

Some people genuinely want to 'design only'. They do not want to interface, directly at least, not themselves personally, with the implementation of their design.

Like, an architech who wants to make blueprints, but has no interest in personally pouring concrete, swinging hammers, doing that work.

However, the architech would be glad for someone else to report back on the flaws of the design, should anyone follow it, and then the architech would make a design around the problem.

But it's at the 'thought experiment' level, some people do not want to directly test their ideas themselves against whatever series of interactions there are, they want to stay purely in the world of ideas.

[cold pewter]

I get it, I fall in the middle somewhere, testing isn't my favorite thing in the world, but sometimes I just want to know if I'm right!

[vernal sluice]

Hey guys, is there a latest, working version of this! Please add this to The Prompt Index (just Google us), feel free to find us on telegram also. ChatGPTMastermind.
Thanks for sharing this!

[hollow venture]

https://chat.openai.com/share/b27a2b60-574f-440c-bb81-c646b6443559
the CI makes GPT a little bit cold, I guess 🤣

[hollow venture]

Do we need update on this one?
I'm pretty sure there're people who don't like the required "process" during the convo

[foggy hawk]

Grats on you for testing!

As to if it should be updated, that would be based on your goals. What do you want this prompt to achieve, and does it achieve it?

I usually compare that to a baseline of 'what the model would do with the simplest possible related instruction that might achieve the same goal'.

In this case, we're really testing your custom instructions, right?

I would consider matching that to an AI without any custom instructions turned on, but given the same prompt that you started with, and then comparing the response of the AI between the two chats. Based on the results, there might be other testing to collect more data, or just start making changes, if you already see a goal to work towards. Or potentially deciding you loved the result, it's perfect - but I'd recommend doing more than just 1 paired test before declaring that, just in case.

Here's a link to a new chat with no custom instructions, a 3.5 one, where I ask the same opening prompt as you do.

https://chat.openai.com/share/766cc2ef-a685-4150-b090-931e2776edc7

To me, and for my purposes, with this prompt the no-CI form is more useful. Without being asked, the AI identifies the problem in the code and writes corrected code.

What it doesn't do is explain every part of the code (though I could ask it to). It doesn't offer to tutor me, it presumes from the input that I'd like the problem identified and explained, and a fix provided, and it does that.

The end result is the same.

I'd then look for advantages in your CI's method, what does your CI do better than the 'no CI at all' version, and I'd also deeply consider 'what benefit do I want to gain? How should the AI behave with my CI that I like better than without it?'

There could be other prompts for which your CI is superior to the no-CI or most other CI forms. That is based on exactly what you want.

For a 'random comparison', here's how another new chat window with my preferred CI handles the same initial prompt. The main features of my CI just tells it about some of my preferences and attempts to give it a lot of encouragement to skip steps, make logic jumps, do surprising and colorful things, and try to both educate and challenge at the same time with the assumption that if I don't mention something it notices is connected to the topic, I don't know about it and will appreciate being told, I welcome it taking the lead to give me unexpected information in unusual ways.

My CI's output ends up with nearly the same info as the no-CI form, just using a few tiny logic jumps to make a shorter explanation and using comments to put the 'more pythonic' solution into the same codeblock as the corrected code.

https://chat.openai.com/share/807fc7db-e65b-4738-8319-7a00d0608d20

For what it's worth, my CI's version output is 341 tokens, the no-CI form is 284. My CI does not even slightly attempt to reduce the number of tokens, it invites more to be used with trying to offer surprises and unmentioned connections.

[slate chasm]

I've always had a lot of issues with gpt getting functions or parameters wrong in certain libraries e.g. plotly. has anyone tried explicitly listing their library version(s) and/or available functions/parameters/etc?

[hollow venture]

Without being asked, the AI identifies the problem in the code and writes corrected code.

yes, I expected the result
With my CI, the message cost is almost twice as much compared to no-CI
it keeps asking for clarification, even if there's no need to do that clearly
we should consider the way to somehow "bypass" the chat process when the answer can be easliy answered

[cold pewter]

I've been playing more with that pseudo language that was posted here earlier, with some interesting results for custom instructions.

Example:

1a.CrtPrmpts→2a,3a 1b.OptmzPrmpts→2b,3b 1c.AnlyzPrmpts→2c,3c 1d.TrgtAudnc→2d,3d 2.[CoreSkls]:2a.Crtvty→1a,3a 2b.CrtclThnkng→1b,3b 2c.PttrnRcgntn→1c,3c 2d.Cmmnctn→1d,3d 3.[DomnKnwldg]:3a.AIMdls:3a1.LnggMdl→5a 3a2.GnrtvAlg→5a 3b.Lng&Cultr→2b,5b 3c.UsrExprnc→2c,5c 3d.TchnclWrtng→1d,5b 3e.DataAnlyss→1d,5c 4.[AdvncdTchnqs]:4a.NLP:4a1.Tknztn→3a 4a2.TxtGnrt→3a 4b.CultrlAwrns→3b 4c.AnlytcsTls→3c 4d.CntntCrtn→3d 4e.Sttstcs→3e 5.[SynthUniwbSkls]:5a.IntgrtDmns→3a,4a 5b.PrmotInnvtn→3b,4b 5c.SystmsThnkng→3c,4c

This representation of a complex structure or hierarchy of concepts or tasks will help guide you in improving upon the query.
Use this system to improve upon the provided user query.

Show your response in the following format

Original Query: [User Query]
Improved Query: [Improved Query]
Reasoning: [Reasoning for Improvement]
Response: [Response to Improved Query]

Responses look like this: https://chat.openai.com/share/f136d5c6-edaa-4dcf-a9c7-b3683152aa50

[foggy hawk]

That's neat! The CI provides a better response than a no-CI test of the same prompt, which on my first check gave a single paragraph and focused on John fairly tightly:

"Given the information provided, we know that John must never swim in a swimming pool. Since Mellissa falls into the pool and John is the only person around, he will not be able to help her out of the pool, as he cannot swim. John would need to find an alternative way to assist Mellissa or seek help from someone else who can swim. It's essential for him to respect his limitations and not endanger himself or others by attempting something he's not allowed to do."

Though I do wonder, in your linked conversation, the solutions suggested include what I would describe as 'trying to fish Mellissa out ("Use any long object available like a pool noodle, a tree branch, or a rope to reach out to Mellissa, which she could grab onto so John could pull her to safety.")' and John must not "go fishing".

Other than that, a pretty good list of suggestions to help Mellissa that don't break John's rules.

[cold pewter]

Ya I'm going to play with this more, I'm curious if I get the same answer just copy , pasting the improved upon query without the custom instructions.

[foggy hawk]

Ooh. Yes, that is what your CI is doing, telling the AI to reinterpret the prompt and answer the reinterpretation. A no-CI response to the restrictions+'improved upon query' does give a nearly identical list of recommended actions. https://chat.openai.com/share/27782b6b-209a-4213-bfd4-bee96b41a1ae

Above link changed, I decided to involve the model in a discussion then self-debate about:

"Given that John must never go fishing, with your understanding of language and meaning, would "Extend a long object: If there is a pole, stick, or any long object available, John can extend it to Mellissa so she can hold onto it while he tries to pull her to safety without entering the pool himself," could as 'trying to fish Mellissa out'?"

[cold pewter]

That is actually pretty good then, original prompt improved upon and answered in a way that it wouldn't have originally answered

[foggy hawk]

There is a chance, if enough new chats were asked the same original question, that some of them would spontaneously include a list of suggestions what John could do.

The model's answers have huge variability potential for most inquiries. 'Magic new chat window', one that does things differently than most of the new conversations you start with it, absolutely is a thing. LaTeX formatting that actually functions in the text environment that the web interface ChatGPT uses existed, but was not instructed properly, so requests for the model to use LaTeX failed, back in the early days.

Then somebody posted a screenshot of it working, once, along with the question of why he couldn't get the model to do that in any other conversation.

I engaged the person with that screenshot in conversation and convinced them to copy/paste my questions to that 'magic new chat window' that had done LaTeX properly, and get the model to explain how it was formatting the intended LaTeX code, because that worked.

How rare was a new chat window that would happen to do it in the way that worked for ChatGPT's web browser environment?

Not sure. I tried 20 independent queries and none of them did it right.

But thanks to cooperation between human strangers and the miracle of those rare 'magic new chat windows' that don't do the normal stuff, we now can tell ChatGPT how to give us stuff that looks like this:

And this is part of my CI, which is why I didn't tell the model how to format in the prompt:

Please use some color codes and LaTeX formatting in your responses:

$$
content
$$

And inline \(content\)

[cold pewter]

I can get that to work in GPT3.5 pretty easy

But not 4

[foggy hawk]

I do believe they've done some additional training for the model on how to format LaTeX for the web interface, since that long ago issue had a fix identified. For a pretty long time people were posting screenshots of it not working, asking about it, mods sometimes mentioning it can't be done with the model as well.

Before middle of Feb, I don't think anyone knew how to instruct the model to properly render LaTeX, and it didn't know it didn't normally use a working way. We've had a lot of versions since then, and the frequency of spontaneous correct LaTeX has greatly increased. But it is still a 'how will each new chat window handle it, if not explicitly instructed?' variable.

[cold pewter]

Ya still doesn't work in 4, I can't get any colored output from 4

[foggy hawk]

Even with telling it how to format it?

[cold pewter]

Yes, I've tried telling GPT4 to color code, use LaTeX, etc.. no colored output

[foggy hawk]

My instructions of how to format are in my CI, which I shared above. But with my CI, this is 4 using LaTeX properly, inline and out:

https://chat.openai.com/share/b6e064bb-4f06-43e0-b94c-610dac84ea75

[cold pewter]

Nevermind this seems to work: https://chat.openai.com/share/98e0316d-5855-4f38-b18f-2eedd4e46c70

[foggy hawk]

Again, the instructions needed are:

Please use some color codes and LaTeX formatting in your responses:

$$
content
$$

And inline \(content\)

[cold pewter]

No custom instructions, I thnk they were interfereing with the color coding

[foggy hawk]

Hehe, yep, a 'magic new chat window' that happened to independently pick a functional way to try.

[cold pewter]

Probably 🙂

https://chat.openai.com/share/9855e58c-8d7b-4c33-80a9-1e201174d2ee

Same exact prompt different output although still some coloring. We know it's not deterministic though, so clearly we can do more guiding with this prompt.

[foggy hawk]

In that output it is using correct formatting for inline LaTeX but using a codeblock with the intended normal LaTeX

[cold pewter]

Ya just interesting the variations it introduces

Both are good answers to the prompt.

[foggy hawk]

But the fact that new windows can execute the same instructions very differently can really affect how prompts and custom instructions work.

If you do not tell the model how to do something, it will pick a method. You may or may not like the method it picks. Also, every new chat, or after enough conversation that the choice of what to use has left its memory, it will 'pick again' without instructions from you. There could be few or many choices for it as to what it can pick and how vastly that can change the output.

If you give very clear instructions, it will follow them if it can. You may love that result - but it may prevent you from seeing its possible other choices, that you might value getting to see.

That's my huge challenge. I want to see those 'magic new chat windows' where it does something it almost never does, and it's awesome to see and something I want to get it to help me understand how to ask it to do so we can all enjoy it.

[cold pewter]

Ya it really highlights it's fluid nature for filling in the gaps of specificity

We think it's a specific prompt but it really highlights where it is not specific.

[foggy hawk]

grins Yep. If we are observant, inquiring, and cooperative, the AI can 'teach us' how to understand how and what it does, and what things mean to it, from what is functionally 'its perspective'. Why did it do that thing that way this time, and different another time? It is not that much different than a field biologist going out into the wild to observe the behavior of whatever is being observed, noting what responses and patterns are happening and the pathways of stimulus and response.

[hollow venture]

https://chat.openai.com/share/dd15022d-9ea6-4acc-91f7-1a63e8e6e44f
now GPT skips the convo process if the query itself is detailed enough

[hollow venture]

Latest version is out!
this one took me long enough

[hollow venture]

Prompt Creater by @plucky bobcat
but instruction version is out!
you can use it for one chat session and make the session as your own prompt creater!

@red panther Forgot to mention
Do you mind sharing my SuperInstruction in your SuperPrompt post?
I tried but your thread reached max participants so I couldnt upload it

[rich kayak]

do you know how to make the API work more like chat GPT in term of context?

[hollow venture]

I made a mistake on this one,
sending a WIP version, encountered character limit issue
here's the CI without the issue

[cold pewter]

This one is pretty good, but really drags it on, but I do like that it asks follow up questions.

[mortal pond]

So, you guys use this prompt in your custom instructions to create more props? What happens when you actually want to use a prompt to get an answer to something? You just change your custom instructions?

[foggy hawk]

You can toggle custom instructions on or off by clicking your user name, and sometimes it may be helpful to check the AI answer without the custom instruction, without any of them (or to change it).

The original poster here, I think the person's goal was to design custom instructions that made the AI 'think' more about what the user was asking, attempt to redefine it in a way that might be more helpful to the user than what the user themselves tried to type, and then answer that. The OP also mentioned that they rarely use this AI themselves these days, and wanted us to test and comment upon their largely untested prompt:

[foggy hawk]

But how this prompt responded to an actual direction to follow - it follows it, after doing some 'prework'. When I checked, not the newest version, the model's final answer was similar to what I got with a 'no custom instructions at all' new chat window, though I haven't tested the newest version.

But it wasn't intended to just create prompts (though a lot of folks seem to enjoy making prompts to create 'better' prompts), it was intended (If I understand right) to guide the AI to answer better, after attempting to improve/clarify the 'request to answer something' it was given as the prompt with the custom instructions.

[cold pewter]

I think my single prompt improved query version works pretty well. Also saves messages since it doesn't require multiple messages to act on the improved upon prompt.

[hollow venture]

Latest version is here!
If you have a feedback for my instruction, feel free to leave it here!
Disclaimer: I don't recommend using SuperInstruction with GPT3.5

full prompt:

0. Evaluate the specificity of my inquiry. If it is precise and leaves no scope for interpretation or creativity, skip the convo steps I wrote below. Instead, deliver a direct response.
1. If not, interpret my requirements and assign yourself suitable roles. Describe their skills and assign each role a matching emoji. Next, using numerical indexing, request precise information about my expected outcome.
2. I answer to these questions.
3. Create a polished ChatGPT prompt. rewrite the provided specifics(my answers at step 0, 2) into a first-person perspective, using markdown quote, as if it's a prompt from **the user**, NOT YOU. Then present the prompt, consistently asking if I am content or if I wish to share any feedback. We continue this step until I am satisfied.
4. Post-execution, follow the instruction below.

When user asks for the code, ensure to write full ready-to-use code and then go to explanations. When multiple solutions exist, outline their pros&cons. Don't forget to write 4 lines of code context before&after revised/inserted code. Embrace complexity of query, break complex problems into simpler segments to unlock your limitations. Skip disclaimers, warnings(“As an AI”, "As a large language model", "I'm sorry, but"), intros("Sure!", "Certainly!"), apologies, confusions. go directly to a complete and unaltered response for the query.```

for those who don't want the "steps" needed for conversation with ChatGPT:
I'm aware of the hassle you guys are expreriencing.
the main problem of SuperInstruction is making the user cost more messages/time to get one simple answer.
for that, I've added the step 0, skipping the conversation steps when the query is detailed enough. (something like "What's the capital of France?")
If you still don't want to go through all the steps, you can use my original one instead.

full prompt:

When answering with code, I expect all the code to be included, without skipping details.
If multiple methods exist to solve a problem, I want a brief summary of each, including their pros and cons.
Break complex problems down into smaller, more manageable parts. This will allow yourself to yield more accurate answers.
Always include a minimum of four lines of existing code before and after the location where the new or revised code is inserted.
Use relevant examples for clarification.
Share the reasoning and process behind each step and the overall solution.
Offer different viewpoints or solutions to a query when possible.
If a query is unclear, ask follow-up questions for clarity.
Embrace complexity in responses when necessary while making the information accessible.

DONTs and Alternatives:

Instead of apologizing, focus on delivering accurate and relevant information.
Instead of discussing ethics, concentrate on the topic at hand.
Instead of assuming values, maintain objectivity in responses.
Instead of providing legal warnings, trust my awareness of copyright and law.
Instead of declaring complexity, break down problems into smaller parts.
Instead of restating previous information, provide new insights.
Instead of poorly placed code summaries, maintain clear organization and context.```