# Tacos

![Across the table at Amaya's Taco, taken just as the recording started](photos/photo-3342a79d-dd17-4901-84b2-8fe61b842335.jpg)

*Recorded 2026-09-25, 38m 41s. A lunch conversation between Chipp and David Avalon.*

## Lunch at Amaya's Taco

We're having lunch today with David Avalon at Amaya's Taco.

The ordering, as it happened at the table:

- One of us wanted one taco and asked whether to get it with beef or chicken. Mixing them was suggested, though it "kind of tastes weird"; they're good, though. Beef was the choice, with no sour cream (not sure if it comes with sour cream; they come with lettuce and a little chicken).
- An enchilada was ordered too. Asked whether they could all be the same or mixed, the answer was yes, so: beef.
- Later the order came out wrong: "These are chicken. I ordered beef." Something else was missed too, so rather than exchange the plate, we asked for a side of it instead ("he can have these, you can bring him a side").
- The food was very good: "Oh, this is so good."
- At the end: something to go, a to-go banana, and sweet tea (sweet tea and lemonade) to go.

## The Recording Device

David's reaction on seeing it: "Now you make me want to go get one of those, man." It's cool.

I'm still working on the design. If I have Wi-Fi here, I can basically hit a button and it'll do the rest.

- **Taking it apart:** David asked if I'd tried taking it apart to rewire it. No, I don't do that, because it's an optimized system.
- **Power off:** I can't power it off right now. I've got to add that. I thought I could go into the system settings, but there's no power off in the system settings for this, because it's a special Android version.
- **Ports:** There's USB. The slot David asked about isn't for an SD card; it's for a SIM card, if I want to put one in. On the real radio (I had one for a while) you can walk around with it, and it's like your own rock box.
- **Microphone:** The microphone is coming off of this, which surprised David.
- **Hours left:** I'm curious about one thing: I still have 2,264 hours left. I'm going to just leave it on until we get to your house.

### David's Wish: Video

David: If that had video, it would just be killer. Not video the way video is, but the way the photos are, taking a five-second clip. I said I can add that. He wanted it because he's traveling, travel style.

## Why Not Just Use the Phone?

The point of my latest video is that I don't want to walk around with two phones.

David uses Granola on his phone, which he says is the same thing. I looked at that a little, but the problem is there are all kinds of things I want to use my phone for, and I don't want that going on in the background. I want to be able to take my phone and walk over there, or go to the bathroom with it.

### David at Church

David controls Granola from his watch, because he uses it at church all the time. The sermon's going on and he thinks, "I want a deeper analysis later," so he hits it, it records on his phone, and it just works. He can shut it off from the watch.

It blows him away, because where they normally sit the preacher is way over there and it's a crowded room. He isn't holding the phone up; it's just sitting on the chair next to him, and it picks it all up. It's ridiculous. They make the microphones in these devices pretty good now, so I think that's part of it. I'll have to think about that one.

### The Bee Wearable

I also have the Bee device, a wearable that does the same thing. You just wear it, tap the little button, and it starts recording. The mics are in front. (David joked that it's a safety thing he certainly doesn't need.)

David asked if it takes pictures and inserts them as you go. No, and that's why I like this one: it has the idea of pictures along with audio.

## What Happens After Recording

This one has three different AI modes:

1. It automatically uploads to my server, or whatever server you have on your computer. You use Tailscale.
2. It transcribes automatically.
3. It generates three different versions:
   - a synopsis
   - a very, very thorough version
   - an executive summary kind of thing

Those three prompts live in the settings, so you can adjust them any way you want, which is nice.

It then uses my MD Docs workflow, but I've embedded that in the server now, so you don't need MD Docs. It automatically generates the HTML files, they pop up, and you can choose from about five different themes.

I'm using that HTML feature for the daily briefing now. I get an MD and an HTML back, and I styled it. You could have a style for IBM, and you could have daily briefings.

David: "Hey guys, here's the minutes of our meeting." "How did you do that? That was like two minutes. I just told you to finish that meeting."

## David's IBM Thumbnail Engine

David already does this kind of thing and messes up his team, because he builds things, and when he shows them, they're like, "What the heck?"

One of his teammates in Dublin asked, "Do you have a way to automate? We need thumbnails." He said, "Give me this afternoon."

What he fed it:

- the IBM brand and photography guidance, which he helped write
- his Midjourney prompt engine, the "filmic" cinematic look
- all optimized for Adobe Firefly

He had to use Adobe Firefly because he can't render outside of IBM for IBM. He can't use ChatGPT, because IBM has to make sure the images are safe: not rendered from somebody's image that they could get sued for. The good news is Adobe Firefly finally caught up.

He incorporated his filmic looks, optimized them for Firefly, and put them on brand. Now it produces incredible, beautiful images, both thumbnails and graphics. He showed his boss today, and she said, "What the heck? Those are AI images?" They look like photographs, and they look like the IBM photography brand. That's the kind of thing he's doing all the time, especially when someone asks, "Can you…?"

### "All He Did Was Press a Button"

What's funny is that people look at these and say, "All he did was press a button." They don't understand. The same goes for the RSA anime workbook he's calling **Scribe Board**. Make no mistake, there's a lot more to it.

The filmic guide took his film and photography knowledge:

- **Lens choices.** If you shoot a face in portrait, you want between a 40 and an 85 millimeter.
- **Film stock choices.**
- **Lighting techniques.** "I want butterfly lighting." What the heck is butterfly? With butterfly lighting you won't see the shadow of the nose.
- **Cameras.** This one shot with an ARRI, this one with a Fujifilm. He can speak to these because he's physically shot with them. People ask him all the time which camera to buy, and the answer is: it depends.

All that intrinsic knowledge is in there now, and it took a while to put together.

## Making AI Films

### Where I Get Stuck

I've got a couple of scripts we can choose from today. I get stuck in two places:

1. I try to run it all as a command-line thing, so I don't have to worry about the interface and can be deterministic.
2. Then, once that's done, I have so many gates and so many steps that I get lost about where I am.

I don't want a fancy interface. My answer is a skill called **MD Questions**. I tell Claude, "Ask me a bunch of questions about this," and it generates a web page using MD Docs with checkboxes, radio buttons and fields. The chat tools only give you little menus with four or five questions; this page could have 50. I work down the page, check things off and add notes, then hit the **Copy Answers** button at the top and paste it into Claude, and now it has a full understanding. That's a good user interface for me, because I don't have to build an app, and I know I have to go back and forth with Claude, since it never gets things right the first time.

My seven stages start like this:

1. Generate the script from a story.
2. Assets.
3. Generate some concept art first, and choose what I like.
4. From the concept art, generate storyboards and asset boards.

I can get that part, though I always mess up the script along the way. From there, ideally, I choose a storyboard to generate first frames from. The trick is keeping context all the way through: when I'm choosing something, I want the context of why. If I'm looking at a storyboard, I want the scene and shot context, everything I need to know.

### David: Three Films — Written, Shot, Edited

This has happened to David, and to Eric, and now they know from building. First there's the story, and you want to nail it down. But after shooting 52 films, the reality is:

- there's the story you write
- there's the story you shoot
- there's the story you edit

The only film the audience cares about is the one you edited. That's the real film.

That still happens in AI. Say we want to do Mission Avalon. We write a script. Once you start rendering, that's the film you shoot, whether it's Omni, Seedance or whatever (the same with "Aquacan"). "That kind of looks okay." In the middle of it you get cool shots you didn't anticipate, and you tweak things: "I changed my mind. Instead of the robot coming from under the ground, I want him to come from above."

The final stage is editing: "I'm going to add some big sounds." "I should have had a cut back to the kid's face, and then come back." He didn't plan that. So, long story short, he now tries as best he can to think about the final stage, the film you edit, because in the end that's all he needs. Still, it does help to lock the story down as best you can, because it feeds everything after.

### Elements and Prompts

The core of how we build AI, which we both know, is that I need **elements**: images for anything.

- character
- props
- location

Those elements get rendered with motion. In the end I need a video, but it all starts with an image.

You still have the style and storyboards to give you ideas, because you're still playing in the sandbox. You've got the story, but now you're in shooting mode: what do the kids look like? What's the kid going to wear? How many outfits do they need? Do they wear the same thing the whole film? Eric's main character, for example, has 14 outfits, and Eric has logos. But in the end we just need that image.

Once we lock the script (say, working with you and Floyd: "Here's the script, we're pretty good, we took care of all the big hooks, everything that needs to happen, and the payoff"), all David really cares about is the prompts:

- the prompt for the image
- the prompt for the animation

You don't need the rest of what I was describing, the concept art and all that.

I asked where the overall style and movement comes from. David: when Eric produces his final piece, how was it done? "I had a prompt that created that cool image." But that prompt came from all the sandboxing.

**Is the prompt generated by you or by the AI?** Both. It's the same science as the thumbnail engine he built yesterday with his co-worker: the filmic lens and lighting knowledge determines the look. We could both take the same story and render it, and one looks like it was shot in a particular style. "I wanted mine to look like 1970, so I did the research: 1970s film stock was Kodak Vista 100, because that's what they shot on, and they also shot on Cooke lenses, probably a 40-millimeter Cooke." That's why it looks authentic, shot in the same period as Blade Runner.

That's done in the sandboxing. But when he finally gets the script, it has ARRI Reflex, Cooke lens, 40 millimeter, butterfly or whatever lighting. Those become staples; they get locked down. Then all he cares about is the prompt for the image, locked down, plus your asset sheets, character sheets and location sheets. Those are the elements.

### Character Sheets: Four Images

Take the little boy in Mission Avalon. Everything's locked down. Eric and David tested it and narrowed it down: you only need **four images for a character**, and that's optimal:

1. a close-up of the head, cut out
2. profile or quarter profile
3. the back
4. the front

So you have your four images and your prompt. Then comes motion.

### Motion Prompting

That's what Eric does. Eric kills it on motion; he's a motion master and has mastered the prompting, through a lot of heartache, because it doesn't always work. The camera doesn't move the way you want. Now he knows you have to name the camera. He doesn't say "Camille moves around"; he says "the camera rotates around Camille." Stuff like that we now know.

So in the end: prompt for image, prompt for motion, done.

### From Story to Shooting Script

I proposed breaking it into maybe three parts, the first being the overall setup: create the script and the asset sheets, but not the prompts yet.

David: it's the story, but the story is written in light of how it's going to be shot. When he writes a film, he's already shooting it in his head. Say we do a short film shot entirely in his car: he'd probably use his 2G M5 and mount some on the outside of the Tesla. He has that knowledge, that's his advantage, and he brings it to AI.

I don't have that knowledge. But David says you have an idea in your head. If he were Claude asking you about Mission Avalon: "Where do you want the camera?" "I don't know, you tell me." "What are you trying to say in the story?" "I want the kid to be very…" "Maybe the camera needs a low angle. Maybe a fisheye lens." That kind of thing helps in the story.

My idea: as you create the script from the story, have a wizard that asks, for each scene or shot, what kind of camera are we using and what effect are you going for? Scary? What's the feeling you're trying to generate?

That's what David does now, and it's why you generate all your prompts in the script:

- **Story Sherpa** works out the beats. It's concerned with the beats first; he calls it the algebra of storytelling, and it has to have certain beats. That's what he means by the story.
- **BoxBox** takes over once the beats are locked. What he's trying to get to as fast as he can is what he calls the **shooting script**. So it goes story, to script, to shooting script, immediately.
- His interpreter interprets the script, which is influenced by Story Sherpa, and gets closer to the goal: the prompts. It speaks prompt language, because that's the only language the screen speaks.

For example, Floyd writes: "The kid pulls up to the thing, frightened, and looks up at the big robot." That's his language. The bot reads it and says, "Camera moves around the element…" It's like English to Spanish: it translates straight away. David calls it the **parser**. So I need BoxBox to generate the prompt script.

### Building Assets

The draft script gets ingested and immediately broken out: so many scenes, so many elements. Elements are one of three kinds: **character, location, prop**. It does the elements first, and immediately converts the characters into parts.

My rule: I only want assets that exist in more than one shot. If an asset doesn't appear in more than one shot, don't make it an asset; we'll just generate it from the style of the movie. If a guy holds a pencil in only one shot, I don't want a pencil asset.

David's logic looks at it differently. If the word "pencil" is used in the script, it matters, because many times objects are important. If you show a gun, why is the camera pointed at a gun? Why mention a gun unless it matters? If it's an object the camera has to look at, he wants detail on it. Eric's character plays her guitar, and Eric knows it has to be a specific guitar. That's a character.

But I pointed out: if I'm sitting at a desk with a light, a mouse and a computer, and another guy sits at another desk, I don't need to specify all that. David agreed: in that sense I'm right, unless it appears in more than one shot, where you don't want the mouse to be one kind in one shot and another kind in the next. The script parsing handles that. But if "Camille looks at the table and sees the R1," it knows that's significant. And "she walks into the diner": you've got to know what the diner looks like, so that's a location. That's also done from the script.

He feeds it scripts in Final Draft format; if a script isn't, he converts it. We won't use the script I sent him, because it has too many computer graphics. I've got another one that already has some concept art we can look at. I'm really interested in seeing how this works and understanding what goes on behind the scenes.

### Batch Everything, or Shot by Shot?

What I was trying to do, and this could be way wrong:

1. Create all the storyboards. Check, check. "No, redo this one." Check.
2. Create all the first frames. Check, check. "Redo this." Check.
3. Render all the images. I'll come back in three hours.
4. Render all the videos.
5. Put them together and swap shots.

Or do you go shot by shot, rather than trying to piece the whole thing together?

David: shot by shot, because you need that art form, the human touch. You can't hit a button and walk away; you have to babysit it, because you've got to direct. You've got to be Scorsese: come in and say, "I need a cut." I said that assumes you're a director. True, but we can now write some of that in, and David has put as much of that knowledge into BoxBox as he can. If you're not going to human-gate it, here's the human input: what are you looking for in this shot? Then it makes sure it hits those things. That's why the shooting script is all he cares about: you can lock all that down in it.

### No More First Frames

Do you even do first frames anymore? No. Sometimes you have to, and you can mix and match, but you don't think in first frames. **Elements are king now**, because now you can do anything. The camera's flying around, you want it up top, and with first frames you'd have to create that frame. A first frame doesn't see anything behind it or in front of it. When it's element-based, it can see behind, in front, left, so the camera can fly around. That's why Eric's motion looks like the room she was in.

Okay, let's go do this. I'm excited.

## Eric Coming Around

Does Eric use skills? No, but he's learning. He used to just type it in everywhere, but he's finally getting it. He's crediting us now ("I listen to you, Chipp"). He just sent David something on Suno, the slider.

The inside joke David posted the other day: "Man, you need to work on that **tortilla skill**." It's a joke about a skill that does nothing. It's a tortilla; it just sits there, with no use other than being a tortilla. (I thought he meant he was going to sell tortillas.)

Eric is coming around because he's now using David's BoxBox center prompt, and he says it saves him a lot of time. David: "I've been trying to get you to. I built it so you can use it, you big dummy." Once Eric locks in, he's good, but he has a mindset that this is the way it's got to be done. The great thing is that David wants him to use it, because he'll use it, break it, and give David 10 ways to improve it, the same way I gave David CD Factory: "Oh, that's cool, I didn't know I could do that."

## T-Shaped Designer

What David loves is that he's taking all of this together. It's called a T-shaped designer: the horizontal goes across the board, and then you can go vertical into anything: photography, design, and so on.

That ties to what he told me before. It's not "jack of all trades, master of none." He's a jack of all trades and a master of all trades. We pinned the exact wording down together: "a jack of all trades and a master of jack of all trades." He doesn't know why he didn't put the T in there.
