Murder Mystery Generator

[simple onyx]

I created a web page for creating a chatgpt3 generated dinner party type murder mystery given number of players and a setting.

The webpage sends the prompt via openai api and returns the information sorted and ready to print and use.

The murderer is 'encrypted' as a qr code. The content is editable directly on the page, in case you want to add some information (such as last names).

I've only tested it a handful of times. I don't guarantee it will work perfectly 100% of the time.

https://murdermysterygenerator-372700.wl.r.appspot.com/

It takes a moment to process. Output, when printed, should look like the attached file.

If you use it and run into issues or have feedback, let me know. I'll try to make improvements. I'm really just a casual coding hobbyist and my skill set is limited. Even still, with my limited know how, I found the openai quickstart helpful and easy to use.

I don't know that I can afford to pay out of my own pocket to keep it free, so if it gets used too much, I'll either have to disable it or somehow start charging for it.

Anyway, I'm excited about what this technology can do and happy to have a place to share my work.

[golden seal]

Awesome

It worked perfectly according to the concept. But there was an issue in that there were only two suspects and one of them was also killed. Maybe that's because I put a low number of players. Heh.

[simple onyx]

Thanks for giving it a try! For whatever number you put in, it should add 1 to it and generate that many suspects, one of which will be the victim. I would say minimum needed for a decent murder mystery would be 4 players, one of which will be the murderer. Might be worth putting that somewhere on the page. Thanks for the feedback!

[pure kelp]

This is really cool. Thanks for sharing this! Do the mystery clues all work out? Like are the games internally consistent?

[simple onyx]

From what I can tell, it does a pretty good job with consistency. Everyone has a motive, so all of the clues can make anyone look guilty (as requested in the prompt that I use), but, for the murderer, the clues connect them to the crime a little more closely.

For example, one scenario I generated was set at a bowling alley. The victim was found with a bowling pin in their back. Which is hilarious enough. One of the clues connected to the murderer was that they were found with a broken bowling pin.

I think part of the fun of this will be trying to figure out why the AI, from the clues given, thought the murderer would be person X as opposed to person Y. I think this happens with human produced murder mysteries too. You can sometimes criticize the clues as being too obvious or too obscure.

Part of the secret sauce too is that I make the request all in one go rather than piece by piece. I think that helps with the consistency. The difficult part was getting the results to be consistent enough to parse back out.

[simple onyx]

Hello. I responded to your post, but forgot to hit reply. My response is above. Thank you for your interest in my work!

[umbral sigil]

Nice work.

[proper obsidian]

Nice, this is exactly what I was hoping to find at some point. I'm totally having a murder party

[simple onyx]

Let me know how it turns out!

[sacred ibex]

Sometimes the clues or other paragraph are set to undefined - maybe you can reproduce it by typing 5 players and "a mansion on a remote Greek island"

[simple onyx]

Thank you for the feedback. I'll look into it. I request chatgpt to format the response a certain way, but it doesn't always follow the directions perfectly.

[ornate token]

@silent barn - High-risk word(s) detected

masonjames's message blocked by AutoMod

We strive to create a positive environment and have implemented additional filters to prevent the use of inappropriate or disruptive language and topics.

Refrain from bypassing the AutoMod filter as per server Rule #10.

[simple onyx]

I gave it a try. Worked for me this time. I would say, if it has an error the first time, you may just need to give it another try. Also be sure to wait a while as it takes several seconds to query and process the data.

Here is the result I got:

[ornate token]

@proper nova - High-risk word(s) detected

BoudaBoy's message blocked by AutoMod

We strive to create a positive environment and have implemented additional filters to prevent the use of inappropriate or disruptive language and topics.

Refrain from bypassing the AutoMod filter as per server Rule #10.

[proper nova]

Hey @simple onyx thanks a ton for this! Just tried it out and love it. Quick question: How do you measure the "performance" of your app? More specifically, how do you know if the output shown to your users is good or bad? Do you use any analytics for this?

[ornate token]

@simple onyx - High-risk word(s) detected

Faullin's message blocked by AutoMod

We strive to create a positive environment and have implemented additional filters to prevent the use of inappropriate or disruptive language and topics.

Refrain from bypassing the AutoMod filter as per server Rule #10.

[simple onyx]

Thank you for the feedback. I do things like this as an enthusiast, so most everything I do is a proof of concept rather than production quality.

I didn't even do the simple task of making sure the data from ChatGPT is formatted correctly, as you can see from the one commenter above that got an 'undefined' error.

While I wanted to create something that could be immediately printed and used, I did use contenteditable to allow users to add or change details to improve consistency or to embellish. So I guess the quick answer is that it is up to the user to ensure the quality of the content.

I would think of this as a casual, just for fun, product made with an ai that might not always be consistent (which can be part of the fun) or as a starting point on which to create a more serious and consistent mm (but would require human editing).

My input to ChatGPT is one very simple request with very specific details about the desired format of the result. I spent most of my time getting that request written well enough that ChatGPT would be consistent in the results it returned. Then I spent time unpackaging those results and formatting them for display on the screen.

My impression, if one wanted more quality control over the details, is that you would have to start with a generic template for a mm and then ask ChatGPT to fill in the details of the template. Your template would guarantee the consistency and the ai would just provide the flavor. I'd be interested to see any projects you are working on or how you are going about making your requests to openai.

[proper nova]

@simple onyx Got it that makes sense! Thanks a ton for sharing 🙂 Does your prompt(s) live in your codebase at the moment?