A couple of GPTs for PAO creation

I made these GPTs to help me make PAO lists. I created separate GPTs instead of one because it starts to get things wrong and forget aspects of the instructions when you’re giving it too many instructions and considerations in one GPT. Separate GPTs also mimics the process of coming up with a PAO list (select people, select objects & actions), so you can use whichever corresponds to where you are in the process. Note that I use categorical PAO (each suit of cards or each ten items has a category eg. Seinfeld characters, sports stars, etc) based on the Major System.

  1. Consonant Character Finder - gives you three people or characters to choose from for each IPA (phonetic alphabet) consonant for the category you provide. It had trouble being consistent with the Major system rules, hence the IPA, but you can find all the major system sounds in the IPA. This also works with non-people (“household appliances” or “emotional states” for example) so it can help fill in gaps in different memory lists. Snippet of output example using “Seinfeld” as the prompt:

  2. Iconic Object Mapper - gives you five different options for an object that’s associated with the person or character. Also gives context for why it chose that object, and suggests three different appropriate actions for each object. You can input a list of people and it’ll give you output for each - I’ve done up to 13 at a time but the quality or amount of context may degrade if you put too many at once. Snippet of example output for Jerry Seinfeld:

  3. Vivid Narrator - writes a humorous story based on your input. I use this to help solidify the person-object and person-action-object in my mind once I’ve built the list. This one was more for fun, but I’ve found it useful in making them memorable. Example story for Jerry Seinfeld and the puffy shirt: “What is this? The puffy shirt ! Look at this thing—it’s not a shirt, it’s a pirate flag ! Who wears this? Nobody! Nobody in modern civilization voluntarily puts this on and thinks, ‘Yeah, I look good.’ But there I was, stuck in it, trapped . You know what’s worse? Everyone else thought it was perfectly fine! Elaine? Nothing. George? Didn’t even blink. Kramer? He’s practically draped in velvet at all times, so of course he loved it. And me? I’m the guy walking around looking like I’m about to commandeer a ship. I mean, what’s next? A tri-corner hat? A peg leg? Don’t even get me started on how impossible it is to eat while wearing this thing. Sleeves everywhere—who designed this?! Pirates couldn’t even eat in this shirt!” [Style: Jerry Seinfeld’s signature observational humor with mockery and incredulity, exaggerated for comedic effect.]

Hope somebody finds these useful, and let me know if any of you have created GPTs like this as well!

5 Likes

I have made several tries to formulate a prompt that would help me find words to complete my 999 PAO list.

I have to say that it really doesn’t work well. Even trying to close all the loopholes so that chatGPT doesn’t make mistakes, it still fails more often than not.

It often doesn’t map the prononciation right (switching digits, missing digits, especially repeating ones, etc.), tends to match numbers wrong (“this word translate to 563 so it’s a fine candidate for 562” (?)), or it doesn’t find interesting words (like “an imaginary word that could mean…” or “here is a word I checked and it’s wrong, do you want me to try again?”)

You can have somewhat good results with a conversation where you point out the errors and redirect several times. But frankly it works just as well simply searching for words myself (or asking other people to help).

In case someone is interested in testing their patience with chatGPT, here’s the one of the prompts I’ve been testing (not the last one, after a while “improving” prompts, you just get worse and worse results).

It was interesting to try, but I’ve pretty much given up on using ChatGPT for this. Either it’s not an appropriate task for it to handle, or I’m just not good enough with formulating prompts.

Please note that I am not using standard major.


You are a linguist with excellent attention to detail. I would like you to help me find words to complete my PAO list.

We are only considering how the words are prononced and NOT how they are written : we are using the International Phonetic Alphabet (IPA).

Consonant Associations: Use only the consonant sounds to determine the numbers, ignoring all vowel sounds. Refer to the following consonant-to-digit associations (using the International Phonetic Alphabet (IPA)):

  • 0 = /r/, /ɾ/, /ɹ/, /ʀ/, /ʁ/
  • 1 = /l/, /ɫ/
  • 2 = /n/, /ɲ/
  • 3 = /m/
  • 4 = /k/, /ɡ/, /ɣ/
  • 5 = /s/, /z/
  • 6 = /ʃ/, /ʒ/, /dʒ/
  • 7 = /t/, /d/, /ð/, /θ/
  • 8 = /b/, /p/
  • 9 = /f/, /v/

The vowel sounds are ignored.

For example : the word LAST, is prononced /ˈlɑːst/. The vowel sound /ɑ/ is ignored. We have the consonant sounds /l/, /s/ and /t/ which means the the number corresponding to LAST is 157.

Because we are working with 3-digits numbers, we only take into account the first three consonant sounds. For example : coffee fairy is prononced /ˈkɒfi ˈfɛəri/, the first three consonant sounds are /k/, /f/ and /f/ meaning that the corresponding number is 499. The last consonant sound /r/ is ignored, because it comes in the fourth place.

It is very important to consider the first three consonant sounds in order. No consonant sound may be skipped or switched with another one. For example : coffee fairy /ˈkɒfi ˈfɛəri/ can NOT mean 490, because there are 2 /f/ sounds and it is NOT permitted to ignore the repeated /f/ sound. Another example : last /ˈlɑːst/ can NOT mean 175, because it is NOT permitted to switch the /s/ sound and the /t/ sound.

Be very precise and check possible answers before giving them.
Example of checking answer :
Starting number : 590
Needed consonant sounds for the starting number : /s/ or /z/ for 5, /f/ or /v/ for 9, /ɹ/ or /r/ for 0
Possible answer : Sephora
Consonant sounds of the answer : Sephora in IPA is /sifôr’ə/, the consonant sounds are /s/ /f/ /r/.
Number corresponding to the consonant sounds of the answer : 590
Comparison of the two numbers : starting number = 590 and number corresponding to the answer = 590.
The numbers are the same, so the answer is correct.
Another example of checking answer :
Starting number : 902
Needed consonant sounds for the starting number : /f/ or /v/ for 9, /ɹ/ or /r/ for 0, and /n/ or /ŋ/ for 2
Possible answer : fairy ring
Consonant sounds of the answer : fairy ring in IPA is /ˈfɛri rɪŋ/, the consonant sounds are /f/, /r/, /r/ and /ŋ/.
According to the rules, we consider the first three : /f/, /r/, /r/.
Number corresponding to the consonant sounds of the answer : 900.
Comparison of the two numbers : starting number = 902 and number corresponding to the answer = 900.
The numbers are NOT the same, so the answer is NOT correct. Need to search for another possible answer.

I think your prompt might have too much information in it. It only has so much “context” it can work with at a time, so it tends to forget parts of your instructions if you give it too much at once. I took a shot at writing a shorter one, which seemed to work for the numbers i gave it!

Help me convert a 3-digit number to a word or phrase, confirming the mapping used and providing 5 options. Each digit will be converted to a consonant sound. Vowels, spaces, and punctuation can be used freely in between, before, or after to create words. Consonants should be in the same order as their corresponding digits and strictly follow the mapping. Extra consonants should not be added before, after, or in between.
Mapping:
• 0 = /r/, /ɾ/, /ɹ/, /ʀ/, /ʁ/
• 1 = /l/, /ɫ/
• 2 = /n/, /ɲ/
• 3 = /m/
• 4 = /k/, /ɡ/, /ɣ/
• 5 = /s/, /z/
• 6 = /ʃ/, /ʒ/, /dʒ/
• 7 = /t/, /d/, /ð/, /θ/
• 8 = /b/, /p/
• 9 = /f/, /v/

Last time I tried the results were dismal.
(Granted that was a few months ago)

What ChatGPT was good at though, was giving alphabetical lists of a category. E.g. musical instruments A to Z, etc.
And to help look for concepts with specific features.

Keep in mind that the LLM can’t “see letters”:

From here:

Language models, such as GPT-3 and GPT-4, do not treat words as a sequence of individual letters. Instead, they break down text into smaller units called “tokens.” Tokens can be as short as one character or as long as an entire word, depending on the design of the model in question and the particular word involved.

My conclusion was that the way they represent the text they “read” and “write” probably makes them unsuitable for to the kind of work we need in transcribing to and from a mnemonic phonetic system well.

Thanks for the Vivid Narrator! It works great for creating memorable stories for learning Attic Greek grammar!