Ep 9: Tightening the Conundrum Desk

0:08Welcome to Curiosity Builds. My name's

0:08Trevan and we are diving right into the

0:11conundrum desk, which is my current

0:13Curiosity Build honoring the Slate

0:15Political Gabfest's 20th anniversary.

0:18It's a podcast I've been listening to

0:19for almost 15 years myself and I'm using

0:22my interest and curiosity about that

0:25podcast to drive my learning here in

0:28public with you. Showing you how I use

0:30cursor, chatbt, claude, different LLMs

0:33and those new tools to code and create

0:36things that I was not able to do myself

0:38even a few months ago. So while we are

0:42diving right in, I am going to step back

0:46and talk a little bit about this and my

0:49vision. We worked through a lot kind of

0:53shaping this into something simpler,

0:55something more easy on the eyes and and

0:58feeling like you can interact with it,

1:00but it's still not where I want it to

1:01be. And part of that is because I'm

1:03trying to make these videos and and show

1:06how I do my process, how I kind of

1:08stumble and I kind of vibe, but also

1:12I've been a little lazy and I've been

1:13rushing things in some ways which make

1:15things slower and I haven't been

1:17applying some of the principles that I

1:19really think need to be active and in

1:22the front of our minds when we're using

1:24these LLMs. I try to, but it's it's a

1:27constant uh it's a constant effort to

1:30build those good habits, not to just be

1:31lazy and rely on the AI outputs to tweak

1:34it for you or to to get you almost there

1:37or to not understand what you're saying,

1:39but you take what they're saying and

1:41just kind of go with it. That's not

1:43exactly what I'm doing here, but it's

1:44not it's not as cohesive and it's not

1:46the vision that I wanted. So, I took

1:48some time to write out what I wanted

1:52myself and I am going to well, I was

1:55going to copy and paste it in here, but

1:57I want you to hear what I came up with.

2:00So, in order to do that, I'm going to

2:03speak to text it, but I did write it

2:05down. Uh, that way I can have something

2:07more clear and then I can understand

2:09where my steps might have fallen apart

2:13in the past. It's different than putting

2:15the agent into plan mode. I kind of put

2:17myself into plan mode. I I sat back and

2:19I thought a little more about what I

2:21wanted the end result to look like and

2:23that's not always what you're doing in a

2:24curiosity build. A lot of times you're

2:26vibing and you're following the path and

2:28most of it is to what we're doing at

2:31Hypandra is to kind of teach and learn,

2:33you know, wow, look at these different

2:34things we can do and how we can

2:35manipulate. Especially if we are

2:37critical and we and we use strategic

2:41application and we question what we're

2:44seeing or what we want to do. We go out

2:46and we look for examples and we ask

2:49other people what they think and we come

2:51back and we use all of that information

2:53to continue to iterate and make crazy

2:55cool things like what I've done so far

2:58with the conundrum desk. It's awesome.

3:00But I want to consolidate that and make

3:04this faster and and do something with

3:08expeditiousness um unlike this long

3:10introduction. And to do that I wrote out

3:13what I wanted. I organized it. I had

3:17chat GPT put it in order for me and then

3:22I am going to use it to prompt. So

3:26without further ado,

3:28I want to rethink the conundrum

3:30generation for the conundrum desk site.

3:32I still like the desk concept and

3:34everything we've built so far. I want to

3:36keep the leaderboard and the voting. The

3:38part that's not working is the current

3:39conundrum generation flow. It feels

3:41clunky. So, let's clean it up and make

3:43it feel simpler and more controllable.

3:46The new idea is that the conundrum

3:47generation should work like a set of

3:49drawers. Each drawer represents a

3:51constraint that the user can optionally

3:53choose. They can open any drawer and

3:56pick the options or leave the drawer

3:57empty. Whatever they leave blank, the

3:59LLM fills in intelligently. The drawers

4:03that I've worked up are as follows.

4:05Drawer one, topic or the what, which is

4:08morals, reality, society, future,

4:10manners, logic. Drawer number two, the

4:13vibe or the setting. Realistic, sci-fi,

4:16magic, absurd. Drawer number three, the

4:19stakes.

4:21Low, medium, high. Draw number four, the

4:24people or the target, strangers,

4:26friends, family, yourself. Jour number

4:29five, the clock or the speed. Right? So

4:33instant, today, in a year, forever, in

4:37the past. If the user leaves everything

4:40blank, it just generates a strong kind

4:42of surprise me conundrum by default. In

4:45general, the output should be tight and

4:47punchy, not essay length, clearly

4:48debatable and not tryh hard, something

4:50people can discuss at a cocktail party

4:54intelligently or comic uh comicably.

4:58I also want to change the voting icons.

5:01Right now they're thumbs up and thumbs

5:02down, but it would be really cool if

5:04they could be cocktail themed, an

5:05upright cocktail, like a martini glass

5:07vibe for an up vote, and a spilled

5:09cocktail for a down vote. If we can't

5:11find the right icons easily, we can try

5:13to do a custom SVG so it looks

5:16consistent. And if we don't have

5:18licensing issu, and so we don't have

5:19licensing issues. Lastly, I want the

5:23profiles drawer so later, if I have

5:25time, I can fill it with links or

5:27profiles for Emily, Jean, and David. I'm

5:30totally fine just leaving that as a

5:32placeholder or coming soon for now.

5:41Okay, so thanks to Whisperflow, it read

5:41everything. I know it looks like I copy

5:42and pasted, but I actually held down a

5:44button and spoke this. And Whisper Flow

5:46is an AI speak to text, so it will

5:48sometimes create nice lists. And you can

5:50go in there and customize it. Uh, all

5:52right. We are going to

5:55give this a try, but I want to put it in

5:57plan mode. Right. So, I put myself in

6:00plan mode. I cleaned it up with an LLM.

6:04I then wrote it again. I just spoke it

6:07in here again. And while I spoke it, I

6:10made an edit um live, right? Um I also

6:15misspoke, so I want to delete that. So,

6:18now there's already been several passes

6:21at this by myself and an LLM. And now

6:27we're gonna have ChatGBT 5.2 Codex put

6:30it into a plan. So that's kind of like

6:32the fourth pass. A lot of passes, but

6:35that's what we want. While this is

6:36planning, I'm going to do some editing

6:38magic so we can arrive at the plan

6:41quickly.

6:43Okay, so I am back with the plan. It

6:46didn't take as long as I thought, but it

6:48did ask me a question that I didn't

6:50record. The question was to fill in. I

6:53forgot for some reason. I'll have to go

6:55back and watch the video. It either

6:56didn't record or I didn't say the people

6:59um that are part of the target over

7:01here. And so those people are friends,

7:03family, strangers, and yourself. So I

7:06answered that question and then it

7:08basically wrote up a plan. Not

7:10basically, it did write up a plan and I

7:12can view the plan here. I can view all

7:14of the thinking that it went through and

7:16then it just delivers it over here in a

7:18way that's easy to digest. The scope and

7:19approach, which I'll read through really

7:21fast. So, replace the current tone and

7:23context generation controls in the

7:24conundrum desk with drawer based

7:26constraint UI while keeping the rest of

7:27the desk. Pass drawer selections through

7:30existing generation client, adding

7:31optional constraint fields for external

7:33LLMs and use them to shape the local

7:35fallback surprise me behavior. We might

7:38have to adjust this external LLMS part.

7:40have to adjust this externaLLM part.

7:40Swap vote emojis for custom martini

7:43spilled cocktail SVGs used consistently

7:46in leaderboard in recent lists. Add a

7:49profiles drawer placeholder coming soon

7:50with no functional links. These are the

7:52files that it's touching that are in the

7:54repo. This is the code UI that it's

7:56going to implement. These are the steps

7:58to do everything that we just talked

8:00about. We're not going to read it all.

8:02These are the to-do lists. It's going to

8:04go through and do all of these. And as

8:07they actually get completed, these get

8:10checked off. If it was a very simple

8:12plan with a couple of to-dos that were

8:14fast, I'd probably show the whole build

8:16and you would just see these tick off.

8:18that I don't need to add time. It's not

8:21very long. It's several minutes, but you

8:24know, we want to get through this video

8:26quick. So, one thing I do want to

8:30change though, I don't want it to go

8:32into this plan yet. I want it to

8:36I'm curious about this LLM thing. So,

8:40actually, maybe we will just have it

8:42build this and then I'll explore that

8:44later. I'm I'm curious what external LLM

8:47endpoint means. I need to doublech check

8:49how I'm set up on the conundrum desk. I

8:52I might need to get an API key to go

8:55into our own API to to access Hypandra.

8:59We'll see. I'll have to talk to Daniel

9:01about that, but off to the races with

9:03the build. It is going to do a bunch of

9:06stuff. And when we come back, we are

9:08going to have a very different looking

9:10site here. So, we'll be back shortly.

9:17Okay, so we're back and it is just about

9:22wrapping up my last little part here. It

9:27is kind of cool. I am very happy that I

9:33took time to write this out and workshop

9:36it a couple of times because still lots

9:39of stuff to do. But look at this. We've

9:43got these martini glasses here. How cool

9:46is that? Um

9:54how cool is that? Also, we've got our

9:54drawers. So, we got the topic drawer.

9:57Don't like how that looks. The vibe

9:59drawer

10:02stakes people clock

10:06profiles coming soon. I think we need to

10:08move that.

10:09Yeah, this is so cool. Um,

10:13all right.

10:15Let's see what happens when we just

10:16click generate a conundrum. When is it

10:18smart to look the other way? All right.

10:21Now, what happens if I just click? So,

10:22let's do reality. Okay. Where does that

10:25show up at?

10:28Oh, it's it holds the whole thing. So, I

10:31think we want this to like it's a it's a

10:34drawer

10:35and it pulls out like a drawer, right?

10:37And you select so then it just sits

10:39there. It doesn't like expand down. But

10:40we didn't weren't that clear. So, we've

10:43got reality, we've got

10:47the topic reality, the vibe realistic. I

10:51think we have some conflict here.

10:58Let's do realistic. Uh the stakes, low

10:58stakes. I like it. The people,

11:01strangers, of course. And we'll do it

11:06today. And we will generate

11:11strangers face low stakes. A reality

11:14vending question in a realistic setting

11:17with the clock set to day. What should

11:19happen?

11:24Hey,

11:24that counts.

11:25Very much counts. What if it does it

11:28again? Nope.

11:30Oh. Oh.

11:33In a realistic setting with the clock

11:35size today. Strangers. All right. But

11:38this is what I was talking about when I

11:40realized I might need to do a consult

11:42with Daniel for doing calls out to an

11:44LLM that actually can provide

11:46interesting conundrums. But I think this

11:49is proof of concept. Very cute and funny

11:51proof of concept. So we are going to

11:55call that an episode and commit this and

11:59work on more changes for the next

12:01episode. So don't forget to commit and

12:03don't forget to stay curious, my

12:05friends.