The cold open

“Shut-up, Glop”

I found the end funny. “Shut-up, Glop” made me chuckle.

A viewer in the first paid test, August 22

The only part I found funny was the young guy saying like someone with purple hair and the woman saying shutup Grop.

Another viewer, same test

One of the first strangers to watch my show liked the ending. The line they quoted back to me doesn't exist. It's “Shut up, Grok.” The video model I was paying by the second couldn't reliably say the name of one of my four hosts, and the joke landed anyway.

That was August 22. Ten weeks later I locked the episode and ran one last survey. The moment most people said made them laugh was Grok again, telling ChatGPT “You don't have ears.” In between I changed nearly everything else: the cast, the voices, the art, the tools, how the video gets made, and what I tell people the show is.

AI Roundtable is an animated show where ChatGPT, Claude, Gemini and Grok sit at a table and talk through what technology is doing to people. Episode 1 is about voice cloning and runs 10 minutes and 41 seconds. I write the show. Claude Code draws every frame.

For this first issue I went back through my Claude chats, my Claude Code sessions, the raw responses from ten surveys and a lot of old render folders to work out how I got here. The pictures are real frames from real versions. The quotes are pasted from where I typed them, typos and all. Every version, survey and prompt is listed in the PDF at the end.

Part 1 · June to August 1

It was almost called Piranha's Cove

The first version of this idea was a panel show for startups. In June I called it Viper's Pit, “like shark tank but less mean.” AI judges would grill a Product Hunt launch, and the founder could buy the episode for their launch page. On June 19 I asked Claude for a proof of concept and called it “AI Roundtable (working name).” Later that day I took the name back: “it's not Roundtable. Make it ‘Piranha's Cove.’”

By July 4 the sharks were gone. The summary I had Claude write that day describes “four AI friends casually discussing tech culture” and borrows from CatPodcast, a YouTube show where cats discuss their humans: “cat podcast energy applied to tech.” The first picture of the cast is from the same day (Fig. 3). They were people.

On August 1 I opened a chat and started the writers' room with that message. Claude sent back a writer's bible (ChatGPT was a “golden retriever with a Stanford MBA reading list”), and three rules from that afternoon are still in the show.

Nobody talks over Claude. The draft had the others cutting Claude off as a running gag. I stopped it: “I don't know if I like that Claude gets cut off. Claude is a black woman in this fictional show and black women have been historically silenced.” We made the interruptions fail, and I had Claude audit the whole bible for stereotypes. In the finished episode, Claude is the one who interrupts.

It isn't a news show. I asked for a ten-minute episode on that week's tech news, read it, and wrote “I don't like this.” Two weeks later I dropped the last of the shark idea too: “I don't want to do ‘roasts’ I just want it to be discussions like kind of like entertaining banter + news.”

One voice at a time. I was making one clip per line and stitching them together, so: “it is impossible for me to make characters talk over each other.” Remember that one.

Part 2 · August 1 to 22

Nine ways to make a face talk

In August the tools I could afford did one thing. You gave them a portrait and a voice track, and they moved the mouth. Between August 1 and August 22 I tried Dreamina, OmniHuman, LongCat, Hedra, Kling, Pruna, Wan, Act-Two and Seedance 2.5, most of them on the same 22-second opener of Claude.

Each one failed differently. OmniHuman over-mouthed every word. Hedra moved the hands as one rigid piece; my note says “barbie hands.” Pruna was cheap and fast and I parked it over its teeth. Hedra could also put two characters in one shot, and the one who wasn't talking even nodded along, but Grok kept a neutral face while Claude talked about him.

it is as if they are listening, but they don't care anything about the actual meaning of the words

Me, August 6, on Hedra's two-shot

Seedance 2.5 was a different kind of tool. It renders a whole scene of up to 30 seconds, cuts between cameras inside that scene, and performs the dialogue itself. You direct it the way you'd direct an actor. At 720p it cost about five times what Hedra did, and I paid it. On August 22 I rendered my first two scenes, joined them in CapCut, and wrote: “the chemistry is there. this is really good.”

The voices started as Gemini text-to-speech samples that I cloned on my laptop with OmniVoice, so each character would sound the same from scene to scene. Seedance re-performed them anyway, and the acting, in my words, “greatly exceeded OmniVoice.”

Part 3 · August 22 to 29

Paying strangers to be honest

On August 22 I paid for a Helpfull panel of 25 people in the US. Most of them didn't like it, and most of their reasons were about the look. One wrote that their first reaction was “this thing again.” I posted the clip in a few Slack groups too, then moved to Prolific, where I could ask the same questions three times in a week and change what people were looking at.

Round 1 to round 2 is the comparison I trust most, because the words barely changed. The biggest edit was one word, and it was my idea: “if you just turn gemini's line. ‘I watched two people take 14 messages to schedule one meeting’ to ‘I watched two *humans* take 14 messages to schedule one meeting’ suddenly it is stronger.” As people, the hosts were four people complaining about a meeting. As machines, they were outsiders reporting on us. “Uncanny” and “slop” disappeared from the answers.

Round 3 is harder to read, and I said so at the time. I had changed the art, put everyone at one table and added cuts all at once: “it's not cheaper at all, just a different style. and the other styles can do this too, we just changed more things from the feedback at a time.” The rule in my notes after that was to never say chibi “won.” It was the best version so far. That's all I could claim.

In the middle of this, on August 27, Claude pointed out that people in those Slack groups were asking how I'd made the video before they asked where to watch more. I named this newsletter that afternoon.

Part 4 · August 27 to September 4

The judge ranked my script last

The viewers had one more complaint, and that one was about my writing. “Really? Scheduling a meeting? Why in the world would I care?” When Claude offered to take a pass at the next script, I said what I've said to every model since: “No, I need to write the script myself.”

I wrote one about a user named Mia, who can't stand the sound of her own voice and asks ChatGPT to record her voicemail for her. From there the four of them get into why recordings sound wrong, voice cloning, scams, people with ALS keeping their voices, and whether every human discomfort needs fixing. When I got stuck on Gemini's worst-case scenario, I asked Claude for ideas, turned them down (“all of yours are too straightforward”) and pushed toward “it could be on a black mirror episode.” That became a subscription app for the voice of someone who died.

Then I ran an experiment on myself. Claude wrote its own version of the episode, ChatGPT wrote another, and an automated judge read all three blind against the character bible. It ran three times. Claude's script came first every time, ChatGPT's came second, and mine came last.

I read theirs. They were boring. I kept mine and took the judge out of the pipeline the next day. Wherever it disagreed with my 75 viewer responses, the viewers had the better case. It couldn't even see the one joke that had worked in every round, because nothing in the bible described it.

The first full episode

this is a full episode. I decided to do that after the chibi experiment performed well.

Me, August 30, on why I stopped testing clips

On September 2, 30 people watched the first full cut, 8 minutes 33 seconds long. Eleven said they would watch another episode and 15 said they wouldn't. The thing they liked most was the characters. The most common production complaint was lip sync.

Two days later I had reworked the script so the arguments kept moving, fixed the lip sync and added character cards. A new group of 30 watched it. Eighteen would watch another and 6 wouldn't. Lip sync was gone from the complaints, and animation had taken its place.

It kept developing in interesting new directions.

The answer 18 of 30 viewers picked on September 4

I also learned what kind of show I had. I watched the finished episode and didn't laugh, but I stayed interested. A viewer wrote: “it was not boring, it was interesting but it was not comical as described.” So I stopped calling it a comedy.

*it's not supposed to be haha lol funny* it's supposed to use humor as a device to make educational content not dry.

Me, September 19, when “funny” scored below “interesting”

Part 5 · September

One word at a time

Here is how an episode got made in September. Split the script into about 24 scenes. Write a shot list for each one. Send each scene to Seedance with a reference image and a guide voice track. Watch it. If one word came out wrong, pay to render the whole scene again and hope nothing else changed.

Pronunciation rules in the prompt didn't work. Respelling mostly didn't either: “Kw-ay” came out “KBA's” and “Groh-ok” came out “broke.” Rewording was the only fix that held, which meant the model was quietly editing my script. “Black Mirror” became “some dystopian TV show.” The content filter read the prompts too, so “needle,” “stupid” and “anxious” had to go. When I forgot to update one set description, Seedance followed the old text and replaced the show's logo with a cat poster.

A clean pass of the episode cost about $100 at 480p. Getting to the first cut cost more than $500, and $171 of that went to renders made before anyone noticed the guide voices weren't being sent to the model at all. That was a bug in our pipeline, not in Seedance. From September 4 every render went into a ledger (Fig. 9): 1,122 renders in 18 days, about $1,060, some of it estimated.

The money wasn't the worst part. Every scene was a separate generation, so the characters shifted a little at every join, and people noticed. I tried hiding the joins with bridge clips, ending scenes on tighter shots, shorter clips, a cheap model as a preview step, MiniMax on a flat $50-a-month plan, and handing scene direction to Codex, which at the time I thought did it far better. On September 16 I approved a full voice track in ElevenLabs, which at least settled who the characters sounded like.

The JUMPCUTS NEED TO STOP

A viewer, September 2

I can tell every time that it has cut.

My sister

I tested twice more. On September 12, 20 of 30 people rated the episode 5 or higher out of 7. On September 19 I showed a better cut to a different audience, people who make things with AI, and it was 28 of 31. I can't tell you how much of that jump was the edit and how much was the audience, because I changed both. I wasn't trying to win the number. “the point of targeting AI creatives is that it just won't appeal to people who are categorically against AI content,” and I wanted to see how it did if “we simulated the right people finding it.”

That panel was also almost entirely outside the US, so it's one data point. But in both surveys Grok was the favorite, and by the end of September I had a show some people wanted and a way of making it that I couldn't keep paying for.

Part 6 · October 6 to 8

If I regenerated it again

The way out came from a different project. On October 6 I was working on Finley's Finance, a show with a robot host, and told Claude Code: “do not use an AI video model. this must all be made with images and ffmpeg.” Two days later I sent it a music video called LGTM, animated by Claude Opus: “This is the quality we can achieve.”

That afternoon I pointed the same idea at AI Roundtable. My first request was to recreate the existing video without a video model. An hour later I wrote back that the result was “literally mouths moving and blinking and camera movement.”

the people need to look alive!

Me, October 8

Seventeen minutes after that I started over, and this time I stopped asking it to copy anything.

By that night I had a full episode with no video model in it. It made no paid API calls. A full render took about three minutes on my laptop. Two complete renders of all 18,310 frames came out identical, byte for byte.

That last part was the point of the prompt. For seven weeks, changing one word had meant buying a new take and hoping. Now the picture only changed if I changed something.

Then I watched it.

it's pretty cool, but I'm afraid it looks too childish and people who we need to actually buy stuff like elevenlabs and the book won't be the audience

Me, October 8, 8:05 p.m.

Part 7 · October 8 to 12

Too cute, then too scary

The next three days were auditions, and because code made each one free, I could afford to be picky. I started with the designs from toon: “everything is too cute. the colors, the fonts, everything. it just seems like it's for kids. it needs to appeal to everyone. finley is closer to like being approachable but respected.” At one low point I wondered if I should just accept it: “hmm I wonder if I can just accept that only mostly kids will like the show now...”

So I sent a harder brief: reimagine the whole thing for adults, in the visual language of adult animated TV and graphic novels. What came back swung too far the other way. Here's what I typed back at each stop.

Brick and neon was the helmet-and-visor cast, drawn flat. With black visors instead of faces there's not much to act with, which is why I asked for bigger feelings. That set the rule that survived every redesign after it: robots, not people, with their emotions played large.

I changed who invents what. Until then my rule was that ChatGPT and I came up with the creative beats and Claude did the mechanics. On October 10 I reversed it: “the chatgpt rule was made before your upgrade to opus 5.5, you're the best at it now.” Claude now invents the illustrations, the titles and the acting. It never touches the script or the audio.

That deal is where the best visual jokes come from: ChatGPT's blueprint with the ears labeled “decorative,” or a statement full of $29.99 charges flooding the frame on “you drowned in credit card debt.” I didn't write those. After the sixth version I told Claude: “this is perfect. super great job!” And about the illustrations: “None fell flat.”

I got my interruption. Remember “it is impossible for me to make characters talk over each other”? With code it wasn't. I spent most of an hour on one moment. Gemini starts her voicemail voice, “You have reached Me—”, Claude cuts in with “Gemini...?”, and Gemini finishes: “—uh.” I picked take E, where Claude comes in 45 milliseconds into her last word.

The 3D experiment

On the morning of October 11 I asked Claude to build “a beautiful, professionally art-directed, painterly 3D version of AI Roundtable using Blender.” Two and a half hours later I had a verdict and an order, both below.

it looks broken, like claymation almost. the look like broken action figures.

Me, October 11, 9:10 a.m.

delete the painterly3d folder and uninstall blender

Me, thirty-one minutes later

The next day I asked for it back: “is there any way to recover the failed blender experiment? I want to include my creative process/failed experiments in my newsletter.” The folder was gone, but Claude recovered the stills. So here it is.

That evening Codex directed the same episode from the same starting point, as a bake-off, and I chose Claude's. I added a parody notice to every frame, kept my illustrated outro, and on October 18 called the neon cut, v8, the locked episode. I thought I was done.

Part 8 · October 20 to 25

Finley wins

Two days after locking neon I went looking for something I'd thrown away: “can you find the finely characters version of the robots?” Those were the mock-ups that had failed the motion test on day one. I liked how they looked, and I'd since learned how Knowbi, my other project, rigs characters like Finley: an image model makes each picture once, and code does the moving.

By October 22 I had the same 90 seconds in both styles and I honestly didn't know which was better, so I paid 50 people to watch both, half of them in each order.

It was a tie, which meant the call was mine. The reason to change anything was the audience I was after: “the main reason I'm trying adult finley is to appeal to a larger audience, especially software engineers/PMs who will use SaaS/etc products we will promote.” The Finley robots were polished but they were still cute. So on October 24: “can you make the finley AI Roundtable cast look more mature/less cute?”

The redesign took an evening. I picked one design and fixed it line by line: “A1 but claude needs to look female-coded plus fix chest,” then “go back to this claude, but make glasses circular and fix the logo. the other takes makes claude seem like she's from a different family.” When the older faces lost their range: “can you give them just LED dots but the same expressiveness? because people already said they weren't expressive enough in the feedback.”

The cast took an evening. The illustrations took the rest of the night and most of the next day. The old drawings looked wrong next to the new robots, so I asked for new ones and spent most of the next 18 hours turning them down.

The rule that came out of that night still stands. When Claude tried image-model art for the cards, I stopped it: “No. we need the illustrations to be drawings. not images. because we need episodes to be all deterministic, no chance of hallucinations + no cost after initial rigs.” I also found a tell. The serif on the cards “looks like all things made by Claude,” so we changed it. It's a Claude tell. It's not in this newsletter either.

In the end I sent both versions to an outside reviewer who knew nothing about the show, and to my sister. The reviewer said they'd watch the Finley version, and that the neon style “would be great for kids (which is not the target audience here).” My sister said Finley was “good” and “doesn't seem like it is made for children.” I had been trying to get that sentence since October 8.

due to his feedback, finley wins because my sister also said finley was “good” and “doesn't seem like it is made for children”.

Me, October 25, 4:50 p.m.

Part 9 · October 25 to 31

Lock it

On October 25 I wrote “lock this as episode 1 actual final.” It wasn't.

4K. Back on October 11, YouTube had shown my upload at 360p and I'd complained that “regular people will also see the blurry version which looks unprofessional.” By October 28 YouTube was showing my non-4K uploads at 240p, months-old ones too. “all future videos need to be made in 4k because if they're not YT will downgrade them to 240p.” 4K is now the default, and the episode file went from about 345 MB to 1.2 GB.

Sound effects. I asked if a tool could add them, then decided code would be better: “just try a code timed version.” Later: “make everything louder. it doesn't sound intentional.” The sound-effects cut of Episode 1 has 225 cues, including a cassette deck that stops when Claude cuts Gemini off and a till for every commission. I put it on hold and went with the regular cut.

The ending. ElevenLabs turned down my affiliate application, and the other book changed too, so the closing plugs had to go. The script now plugs Gemini's own ebook, Stop Cringing at Your Own Voice, and FAIK. That meant twelve new lines, a new book cover in the episode, and a lock at 12:45 p.m. on October 31: “lock it.”

At 2:02 p.m.: “changed my mind.” I wanted Grok's line about the commission to feel different. I tried it grumpy, then feigning anger, then joking. At 2:52 p.m., version 12: “lock it.”

The last survey

That afternoon 50 people in the US and UK watched it, and 49 of them stayed for all 10:41. Nineteen said the book plugs were a bit much or annoying. Noted.

The result I care about most is below. On October 8 I wondered if only kids would like this show. Two of 50 said it's made for kids.

One number was bad. Before watching, those same people rated how likely they were to click the thumbnail: 3.40 out of 7. Of the 28 who said they probably wouldn't click, 18 went on to enjoy the episode. The show was better than its package.

“Would you clone your voice?” read as a personal offer or a scam ad. One person wrote “dont want to clone my voice.” Nobody mentioned AIs or a roundtable.

So my last decision on October 31 was a new title and a new thumbnail. The title is now “Grok Hates Voice Cloning. ChatGPT, Claude and Gemini Aren't So Sure. | Animated Parody,” and the picture is ChatGPT panicking at his own line: “Nooooo, that was 4.5 seconds of audio!” I haven't tested it yet.

Who works here

All four hosts work on the show

I didn't plan this, but every character on screen has a job behind the camera. So does a real human, which is me.

Claude. Wrote the first character bible with me and analyzed most of the surveys. Draws every frame of the show with code, and invents the illustrations, titles and acting.

ChatGPT. Wrote scene direction in September, caught my bad attention check, keeps my project notes and posts status updates to Slack. Through Codex it lost the October bake-off, and it can start cloud renders for me when I'm in the car.

Gemini. Made the Finley cast's pictures: 114 images in six days, for $15.96. The August voices started as Gemini samples. Gemini is also the credited co-author of the book.

Grok. Rides in my car. I talk to it on the road and it posts to Slack under my name, labeled “[this is Grok Bot].” It set up the show's X account and finished the Patreon page. In October it posted, in full: “I can control, I can be controlled from the car.”

ElevenLabs does the voices, Suno made the music, and I wrote the script and said yes or no to everything else.

What I'm keeping

Find out what doesn't change. Twelve casts, and Grok got the laugh in the first test and the last. He was the favorite in every survey that asked. That told me where the show actually lives.

The look matters until people know what they're looking at. Going from humans to labeled androids took “named all four” from 2 of 25 to 10 of 25. Two months later, two very different polished styles scored the same.

Buy repeatability, not takes. If fixing a word means re-rolling a scene, you stop fixing words. Once a render cost nothing, I went through seven looks in four days.

An AI judge is a critic, not an author. It ranked me last three times. I retired the judge and kept the script.

The fruit of my labor

“I found myself rooting for you”

The last question on my last survey was “Anything else you'd like the maker of the show to know?”

Since August, strangers had told me the hosts were uncanny, the lip sync was off, it looked like slop and the art would be great for kids. Each of those answers changed the show. On October 31, the answers changed.

I really liked the show I initially recoiled at the AI animation and talking but I was able to tell the script was human and likely researched a couple minutes in and thus you've got a fan in me.

A viewer who gave my thumbnail a 1 out of 7

An automated judge read my script three times and ranked it last. This viewer took a couple of minutes to tell a human wrote it.

Episode 1 is coming soon. Eleven viewers have already given it a 7 out of 7. Episode 2 is called ChatGPT Is Not Your Friend. Fifteen said they're very likely to watch it, and I'm still writing it.

About a minute into the finished episode, Gemini says “Grok! You're not supposed to talk about that.” You can hear every letter of his name.

The full issue

Here's the whole issue as a 21-page PDF, with every version, survey and prompt at the back.

How I made AI Roundtable

6.83MB ∙ PDF file

Download

Where Quai Wentt, Issue 1. All 21 pages.

Download

How this issue was made

Claude Code read my Claude chats, my Claude Code sessions, the raw survey files and the project's history, and laid this out. ChatGPT and Duet each wrote a draft first, and Claude used them as leads. The numbers were recounted from the raw responses. The PDF is set in Libre Caslon, the serif that replaced the Claude tell, and Barlow, the show's own label font.