Sep 25, 15:59
A working session between two people, walking one test project from a finished script to a first rendered shot. One of them (the filmmaker) built Story Sherpa, BoxBox and CinePrompt and drives the tools; the other is learning the workflow in order to compare it with their own. The recording doesn’t label speakers, so this document names them only where the talk makes it clear.
Purpose of the session
- This is step two. There is already a story and a script from Story Sherpa; now it gets converted into something that can be shot.
- The point is not a perfect ad. It’s to understand the workflow: “The purpose of this is for me to understand the workflow, not to get a perfect ad.”
- Two things are happening at once: capturing the workflow process, and building the piece. Working alone, the filmmaker says they would already have scenes done, because they would have cut through a lot of the discussion.
- Late in the session they agree to split the work into two sessions because time has run on. The aim for this one: get even one scene done, so the workflow makes sense.
The script: too many screens
The concern
The Story Sherpa script leaned heavily on screens: shots of someone looking at a screen. The question was whether that was a problem.
The filmmaker’s view: the script has screens because, in the end, the hero is the software. But the way they visualized it, there is Jordan, the person who personifies the tension, and the script shows the screen just to make a point (“oh, Maria’s email is there”). The rest of the UI they didn’t read that way. It’s more about focusing on his face as he reacts to what’s on screen, because that’s what the audience needs in those 60 seconds.
Can AI make those screens?
- The product itself isn’t shown in the script, so it doesn’t matter. The screen does: the vibe coder is working on his own app.
- The look wanted: those film screens shot so close you see almost every dot glowing, maybe with scan lines, something that communicates it. Part of that comes from the mood board.
Measuring it
They asked the AI what share of the 60-second script is focused on a screen:
- About 50%, roughly 33 of the 60 seconds (it also said about 32).
- The end card takes the last 13 (or 8) seconds.
- Counting the end cards, which are also graphics, about two thirds of the piece is graphics rather than people.
Rebalancing
First answer, rebalancing without adding time:
- Put Jordan in the console section.
- Cut to his face.
- Use shots that show both at once.
- Make the montage more human.
That would land around 40–45% screen and 40–45% human. But that is a compositing effort, which is fine, but not what this test is about. So they asked for less focus on screens, maybe 30%, because they want something more exciting visually.
“A lot of stuff can be fixed in script” — something said at lunch that stuck. Before going all the way to trouble, fix it in script. That is exactly how the filmmaker thinks next: visually.
Three moves (the AI’s revision)
- Give the agent a voice. Right now the agent replies with text, which forces a cut to the screen to read it. If the agent speaks, calm, the cut isn’t needed.
- Turn the scan into a set piece. When NewScan runs, the dark room around Jordan fills with a translucent glowing map of his app.
- Revised beat sheet:
- Cursor. Deploy. Then his face, in screen glow.
- The “Ship it” note.
- A physical montage: his hands, coffee piling up, a time lapse of night and day through the window, the day counter, Jordan grinning and pacing, the agent’s voice with a few quick "done"s.
- The music dies. The “Logged in as Jordan” header, the URL tick and Maria’s face stay on screen.
- A slow dolly zoom on Jordan as the room seems to stretch around him, and his glance at the sticky note.
- He asks out loud, “Are we secure?” The agent’s voice answers. We stay on his face for it.
Both liked giving the agent a voice and the other ideas. Asked whether any of these shots are impossible for AI video, the filmmaker said no, they’re all fine. The other had assumed a split screen or a translucent screen would be impossible; the filmmaker said that’s no problem.
They told it: use the agent voice idea and the rest, write it into the full script as version four, and put it in Final Draft (Hollywood screenwriting) format.
Why Hollywood screenplay format
- It’s the visual language of cinematography. Seen that way, the script reads differently.
- Sherpa doesn’t explicitly talk about Final Draft format, because the filmmaker works with engineers. But any LLM understands it (it may be more a Claude understanding than a Sherpa one).
- The reason it matters: when the script goes into BoxBox, it does something different. It gets one step closer to “the translation”: in the end BoxBox translates from English to Spanish, so to speak — from the client’s story to the shooting language.
- Notes taken for later (no action needed): (1) export scripts as Markdown; (2) understand the concept of Hollywood screenwriting format.
Fountain, Final Draft and FDX
- The script came out as a Fountain file. Final Draft (paid; there are free alternatives) wouldn’t open it.
- The AI checked: the Fountain file is valid. Renaming the extension didn’t help either. Final Draft’s own format is FDX, which turned out to be structured XML, not a proprietary binary. That’s good to know: it can be read, and a reader could be built for it if needed.
- The Fountain viewer only views; it doesn’t edit.
- Once the FDX opened, it was easy to read: Jordan’s apartment, near silence, laptop fan hum, extreme close-up on a laptop screen, the cursor, “Deploy to production.”
What Final Draft adds
- In Production you can add scene numbers. It then knows how many scenes there are: 11 (it counts the cards as scenes). It doesn’t do shot numbers, which the filmmaker would love.
- You can do visual tweaking there if needed, e.g. “Deploy to production, focus racks to Jordan.”
- It shows where dialogue is. Scene one has none. Scene two, the montage: the driving synth builds, Jordan’s hands fly across the keyboard. JORDAN: “Build a login flow.” AGENT (calm, neutral, a little dry): “Done.” Coffee cups multiplying across the desk. JORDAN: “Add social login.” AGENT: “Done.” Time lapse through the window: day, night, day. Whiteboard.
- It has no timing, although the original script did. That’s where it gets interesting.
Where style, look and casting come in
The question: where do the overall style, coloring, lighting and camera choices come in? Where does the treatment go — e.g. Jordan is a scruffy white guy of about 35 with a scruff beard and a baseball cap?
In the filmmaker’s process, that comes at the World step: build the world, aspect ratio, format. It did pull some of that automatically from the script, e.g. 1.85:1, which is a good aspect ratio with a more cinematic feel.
Order of work: style happens first, because you’re trying to lock the look and feel. And characters have to be consistent, so you build the character LoRAs/elements; otherwise you get three different characters.
BoxBox
Versions and how to get it
- The latest BoxBox (BoxBox Studios, a.k.a. the new one) sits in a GitHub repo, and the other person already has access. Every update pushes the latest to the repo.
- It’s just a website; it could be turned into an Electron app. It uses an API.
- The filmmaker’s software tools include a script manager (a whole script-writing tool), a poster maker, and BoxBox:
- Version 1.0 — very simplified, was great and still works great, and taught a lot.
- Version 2.0 — BoxBox CinePrompt / BoxBox Filmmaker, the working one they’ve been using.
- Version 3.0 — BoxBox Studios, the newest, which brings everything in (world, etc.). The last commercial was built in 3.0.
- The older one has some aspects that are better, so the filmmaker tries both. That’s also why Eric likes the older one.
- Eric has just started using BoxBox, and now knows how to tweak it. But Eric took out things he didn’t need that the filmmaker does need — including the character builder sheet. (“Dude, you took out my character builder!” “You can’t let him have access to your source code.”)
- The visitor hopes to talk the filmmaker into moving BoxBox Studios to Electron, or will do it themselves and fork it.
First import (before cleanup)
- New project, named NewsCam (NewScan). Dropped in the Markdown script.
- It broke it down immediately and fast: 30 scenes, 38 shots. That was too many, because the script wasn’t in proper script format.
- There’s a shot list. The breakdown is what the filmmaker checks first: how it read the script.
The scene-count problem
After reimporting the Final Draft version, scene headings still caused trouble:
- “IN MONTAGE” was read as a scene when it should be a transition (a smash cut into scene two).
- “INSERT LAPTOP” and “BACK TO SCENE” lines were read as new locations. “Back to scene” means back to where we were before — the home office, night.
- Headings didn’t match each other; the montage had no location at all, so location extraction would return nothing.
The real point: it’s all one scene in one location, and the script must read that way.
Thinking like a director of photography
Forget AI. If you hire the filmmaker to shoot this, as DP they have to work out how many locations you’re shooting at, and how many shot setups. That’s the homework for the schedule: we shoot these scenes, and they’ll take so long. BoxBox needs the same thinking. It isn’t perfect yet; the filmmaker keeps finding ways to get it to learn.
Breakdown discussion with Claude
They asked Claude to anticipate how BoxBox/CinePrompt will break the script down into core elements — characters, locations, props — and to make sure the script is worded properly. Its answer:
- Jordan — the only on-screen performer, “late 20s, hoodie.” (No hoodie wanted; fix later.) The one real character.
- Agent — voice only, no visual form. The parser creates a character element and expects a reference image, so it should be marked voice-only.
- Narrator — voice-over at the end; same problem as the agent.
- Maria Lopez — a photo on a profile page; she never appears in the room. Her name is in all caps.
- NewScan — the VFX entity; behaves like a character; in all caps.
- Locations — only one real location, Jordan’s home office, in three lighting states: night by screen glow, the NewScan architecture overlay, and dawn. (Neither person knew what the “NewScan architecture overlay as a VFX state of the same room” means.)
- Problems — the scene headings don’t match; the montage has no location; inserts are misread.
Matching the script to the parser
- The filmmaker wanted Claude to read the actual parser code in BoxBox Filmmaker (Library → CinePrompt → components → script parser), so the script fits the parser. Claude couldn’t read the parser code through the browser and asked for files to be pasted in; the filmmaker told it not to, and to focus on the first three scenes instead.
- Aside on Claude permissions: it kept waiting for approval; the fix was switching from manual to allowing more than once, then cancelling and rerunning.
- The filmmaker has done this before piecemeal, bouncing between Claude and CinePrompt. That’s why they want to build Claude (or whatever) into BoxBox.
- The instruction given: make any changes needed to the Final Draft script so the BoxBox CinePrompt / BoxBox Filmmaker parser understands the scenes, characters, locations and props correctly — either rewrite the script so it’s understood, or write code to understand it better, whichever makes more sense.
- Tweaks can still be made inside BoxBox after import.
Result: version five
Claude’s changes:
- Every scene now has exactly the same heading, except the last, which is its own dawn scene.
- Scenes are numbered 1 through 6 with no gaps.
- The montage has a real location heading.
- The three INSERT LAPTOP lines and the cards are now shot lines, not scene headings.
- Jordan matches the casting direction: late 20s–30s, scruffy beard, plain worn baseball cap, no hoodie.
Imported into BoxBox as version five: 16 shots. “That’s much better. We’re back on track.” Later, parsed again in 3.0: 2 scenes, 16 shots, 3 characters. The filmmaker is bringing in things from the previous version (which is why Eric likes it), plus the cleaned-up script.
The open problem for the filmmaker: how to get from a Story Sherpa script to a Final Draft script that BoxBox can use. Right now that happens by hand through Claude. Once the story is nailed, what’s left is the shooting script and the Midjourney/Kling prompts — “that’s what I haven’t perfected yet,” and it’s what should be as fast as possible.
CinePrompt
- CinePrompt is the prompting engine, part of BoxBox. For the other person it may be the key part to grab from the repo, to see how it works. (“It’s a good name, by the way.”)
- What it does is the translation. The script in Final Draft is natural language that clients can read (“oh, I can visualize him at the screen, looking”). That satisfies the client’s need to understand the story and sign off. Then the real work begins — the equivalent of going out and shooting it, except the cinematographer is Midjourney and so on. So it translates into the shooting script.
- Its point is to take a script and convert it into prompts for the video (and images), not for the objects. Get to the prompting phase as soon as possible, because, as with film — you write it, you shoot it, you edit it, and the edit is the real one — that’s where things really happen.
- Is it an AI? Not at heart. It has deterministic components, but a lot of its rules were written by an AI. It knows, for example, that if the imagery comes from Midjourney, it must speak Midjourney’s language.
- How does it know the character’s age or the film’s style? That needs to be supplied — sometimes in the script, or by giving it more information.
- It generated multi-shot prompts from the script, following the shot list, using the camera material the filmmaker had sent over. Camera choices come from the project’s style, which wasn’t set up yet. The film style section has keywords, ratio and style images.
- CinePrompt runs on the visitor’s own LLM too. But in BoxBox Filmmaker, LTX isn’t plugged in yet: the filmmaker ran the tests in Claude but hasn’t coded the video part. The prompts can still be pulled out, so both approaches can be compared. (Tokens aren’t an issue: unlimited.)
Finding a cinematic style
Reference research: ShotDeck
- This is where the art comes in. The filmmaker uses ShotDeck (discussed before) to see how others shot things.
- Searched “computer screen”: You’ve Got Mail (kind of cool, not the right feel), Moneyball, The Matrix (“feels pretty good” — the kind of cinematic feel wanted).
- Also looked at existing concepts on the hard drive (Anthony → concepts → screen plates v1). These are Midjourney images already made for this piece. Some characters liked, some not. They have a mood but not a cinematic feel. “It’s got a mood, but it’s not there.”
What “not cinematic” means
The visitor thought one image was a beautiful rendering. The filmmaker’s view: it’s cool but not cinematic. The lighting is burnt, and he looks drawn, like an illustration. It isn’t filmic; it doesn’t look like a photograph, even though the prompt asked for one. That difference was new to the visitor.
The filmic prompt generator
- The filmmaker keeps tools in ChatGPT, including a Midjourney prompt creation project and a “filmic cinematic prompt generator.”
- Workflow: take an existing prompt and ask the generator to rewrite it for a cinematic look and feel. Keep it simple.
- It started generating an image instead of writing a prompt; they stopped it.
- It added “AI Story Sherpa” to the prompt although Sherpa was never mentioned — it admitted the mistake. Also removed “no text, no watermark” style lines; the filmmaker often doesn’t need the “no” stuff.
- The prompt was too long. Midjourney 8.2 has an optimal prompt length, so they asked it to rewrite for that. After that, generation ran much faster.
Midjourney settings used
- Aspect ratio: 16:9 was considered; went to 2:1 for a cinematic look.
- Weirdness 0, variety 10, stylization 10.
- Version 8.2.
- The filmmaker’s default profile, tuned over cinematic images. A mood board could also be given; this run was without one.
- Casting added: engineer mid-20s to early 30s, male, Caucasian, baseball hat, a little scruffy beard, not a hoodie.
The rewritten prompt (read aloud)
Software engineer reading an open laptop at 2:11 a.m., authentic Caucasian male, age 28–32, short scruffy beard, plain worn baseball cap, fitted dark charcoal Henley, thoughtful expression, hands resting on the keyboard, quiet concentration, minimalist wooden desk, blinds closed, lamps off, laptop screen is the only light source, casting cold blue-white illumination across his face, hands and desk, crisp photorealistic skin and wood detail, premium cinematic office, eye level, oblique three-quarter angle, asymmetrical composition, medium shot, motivated screen-light lighting, low key, cinematic shadows, David Fincher, cool cinematic palette.
Results looked good. One has a hat; the filmmaker liked one the most: good style, and a good face for the character. The generator link was shared by text.
Critique of earlier concept images
- Too centered, which is why “asymmetrical” in the prompt is a good point.
- He looks sad.
- The characters feel too muscly. Developers are skinny dudes (“whatever that is — me”).
- You’re mainly looking for color and lighting at this stage.
Mood board and film style
- The filmmaker searched ShotDeck for screen looks: “hacker screen,” “data screen,” “CRT screen,” “digital display” (an American history one could work as a close-up, old-school), “sci-fi screen.” “Data screen” gave most of the good ones.
- A mood board image influences the look. The risk: use a specific image and it will copy it, which isn’t wanted. You want images with the right color.
- The film style is a prompt influence: it shapes any prompt the tool generates, for any image. So even when generating with Kling, it should affect it.
- There are no storyboards and no establishing shots in this flow. Film style takes effect when the tool writes prompts in the Shot Builder.
Shot Builder and multi-shot prompts
- Scene one: Jordan’s apartment, near silence, a laptop fan hum, extreme close-up on the laptop screen. Within a scene come shots 1.1, 1.2, and so on.
- “Generate the multi-shot” writes a prompt for Kling 3.0 multi-shot, in exactly the form Higgsfield wants, so it can be pasted into Higgsfield. It produced three multi-shots. Example: “…the camera dollies in a dark cluttered home office at night.” “That’s what I want. There’s the shot.”
- The prompt includes camera, shot list, setup (Jordan, apartment, desk), and movement (pull in, push). Some came from the script: it chose an extreme wide shot. The filmmaker also changed things, e.g. a Dutch angle, which “will be an interesting shot”; there is also a “superhero / graphic novel” style option.
- Element references use @ names. Clicking @Jordan shows the element; if an element doesn’t exist yet you create it. References are matched by the image name. Red means unmatched; once fixed, both turned green (a misspelling of “apartment” had to be corrected).
- The metadata block at the top (shot 1A, etc.) — whether Kling wants it was left open.
Building the elements
Characters
- Before rendering, the elements have to be built. In the character element: name Jordan (as in the script), a description, and at least one reference image.
- One idea was to use the filmmaker as the developer character, since they’ll need a character of themselves for the podcast anyway. They shot reference angles: front on, semi-profile, back, and face forward.
- The character builder was entered with: age 22, Caucasian, wardrobe baseball cap and T-shirt, “kind of rough and cool,” vibes programmer, live coding, slim, scruffy.
- There is an “Enrich with AI” option.
- Since Eric had removed the character builder sheet, the filmmaker pulled the stored character sheet prompt (“character image builder”) and gave it to ChatGPT with version 4.5 of the script: “There is a character named Jordan; build the character sheet for Jordan based on the script.” It did so without a reference image, because it had the script. (That skips the usual step of generating four candidate characters and choosing one; the real process would build a couple of characters and decide.) Claude can’t generate images, so this ran in ChatGPT.
- The result: “There’s our character. Save him.” Then the same for Maria — only her face is needed, but a full sheet was made. It did only the face, which was what was wanted.
- They then asked ChatGPT to cut the Jordan sheet into seven separate images for upload, because that is how a LoRA/element gets built in Kling, which doesn’t use first frames here. Then the same for Maria.
Kling versus other models on character sheets
- The visitor pointed out that Grok and Seedance can take a full character sheet and keep the character the same in every scene. The filmmaker said Kling can’t; it needs separate images.
- A new Kling version is supposedly due any day. Both the filmmaker and Eric noted it hasn’t been updated in forever. The creator of the “flame flower girl” videos is also doing the Kling announcement.
- Kling is one of the cheaper models. The filmmaker has credits; some videos are free. The visitor thinks they used a character sheet with Kling once.
- In Kling’s element upload there seemed to be only two upload slots, and an aspect ratio error for the character sheet. The visitor won’t likely use Kling anyway; they’ve already solved character consistency their own way on their last video.
- Voices are added in Kling, or with Grok or H3.
Location: Jordan’s apartment
- Locations were generated with the filmic Midjourney generator reading the script. It gave several, including one with warm morning light. The high-tech one was liked.
- Images can be adjusted in Higgsfield or ChatGPT; ChatGPT image is now pretty good and starting to compete.
- Edits asked of ChatGPT:
- Make the office a modern high-rise apartment in a big city (Austin), with blinds and clean lines.
- Keep the teal and orange cinematic color grade.
- Keep it dark and foreboding: the opening scene needs tension, so shift the night colors for more drama. Keep the city light outside the window, but only one light at the desk — no extra lighting around the desk — and a bright monitor.
- Saved as the location “Jordan’s apartment night.” Then a dawn version: the same image, bright and happy, as the script’s ending calls for. The yellow “Ship it” sticky note shows on the laptop.
- The dawn image still didn’t look cinematic, but that will be fixed.
Making images more photoreal
- Idea: ask ChatGPT for the prompt that made the image, run it through the Midjourney filmic generator, and use the image as a reference. A possible workflow: use one tool to get photorealism, then use the result as a reference image (an omni reference for the videos, or in the mood board).
- The Midjourney result changed the image and wasn’t as good. Better: keep the regular two images and, in the video prompts, say “this is the room; make it cinematic, depth of field.”
- “Shot on RED” in the prompt looked like part of the problem. Suggestion: take the characters out in ChatGPT and try “shot on RED” there.
Putting Jordan in the room
- In Nano Banana Pro, you can give the full character sheet with the night location and say: add the character to the chair, sitting at the desk, facing the laptop. It worked well, using the full character sheet.
- This isn’t an establishing shot as such; it’s an omni reference, a reference for Kling. It could serve as an establishing shot, and from it the filmmaker would generate multiple views.
- “Turn to video” from there would use a first frame, which may not match what the story wants (a foreboding opening).
First render
- Back in BoxBox 3.0, reparsed (2 scenes, 16 shots, 3 characters). Characters: agent, Jordan (software developer), with a reference image uploaded. Location: Jordan apartment night. Laptop element not needed yet; a wall clock reference was left off for now.
- Pasted the prompt into Kling 3.0 Omni. Multi-shot means it breaks the shot into multiple cuts. At one point switching modes deleted the prompt (“you dirty rat”) and it had to be rebuilt.
- Shot two would be: smooth lens, zoom in on laptop, screen reveals a profile page for Maria Lopez with her home address visible, “Logged in as Jordan.” Here a laptop element would really be needed to control the screen.
- Decided to render just the first piece as a single shot, not multi-shot: 8 seconds (10 was also mentioned), 720p rather than 4K, sound available.
- Then the same prompt went to the open-source model, LTX 2.5 (installed on the visitor’s machine), via Claude in the terminal, with the elements as references. (“You should make a skill for that.”) The Kling render took almost no time. LTX generated a first frame — “so it does want a first frame.”
- Is the result cinematic? Yes: a cinematic move, and the image is pretty good. “In the end, that’s all I care about.”
First frames, omni references and camera moves
- The filmmaker doesn’t use first frames. With first frames, you can’t pull off the bigger cinematic moves of modern filmmaking — circling a character, through the car and out into the city, like Spider-Man. A first frame doesn’t let the model see what’s behind you and in front of you, so you get drift, location drift.
- The visitor’s counterpoint, from a video they watched: with a first frame plus an omni reference for the location, the location holds. Their location sheets show the room from every angle, so as the camera spins, the model can see the rest.
- The filmmaker’s reply: feeding it a multi-view image is like your own version of a LoRA — though technically a LoRA is a different thing (training, “a pain in the ass”). What they’re really talking about is omni references.
- The filmmaker uses omni references on everything, but only the ones a shot needs. If a single shot has him walk through a foyer with a tree on the left and a key jar on the right into his office, they never use the foyer as an omni reference unless he comes back out. It can render whatever; it isn’t important. A pen is different: if he drops the pen on the table and stares at it, it matters.
Finishing: the color grade
- The last layer is the final color grade, done in Premiere or Final Cut. Both the filmmaker and Eric do this.
- The filmmaker’s tool is called Film Emotion (or similar). It broke on this Mac after an Apple update and needs an update — “I’m so bummed,” because it has all the presets set up.
- Always 24 frames per second, for that film feel.
- Applying a LUT (CLUT) adjusts the colors; film noir is an option.
- The visitor’s usual approach for a video look: add some film grain, pump up the contrast, pump up color intensity, burn them out a little. That’s what a lot of LUTs do. Also lens glare (not lens flare), and a little vignetting.
- Blender’s compositor also does this very well: a full compositor, “out of this world.”
Takeaways (the visitor’s summary)
- The Story Sherpa work was incredibly valuable.
- Also valuable: you go shot by shot by shot, not all at once. The visitor had been trying to do everything at once.
- Their own process mostly tracks this all the way up to video: they have all the characters in files. But they probably spend too much time on storyboards: a storyboard for every shot, refined into an initial shot, used as the video’s start frame. That’s roughly what this does too.
- What they lack is CinePrompt, now understood as an important part, the way to short-circuit the whole thing. The catch: many of the models they use need a first frame, so they aren’t sure they can use CinePrompt without one. Their system currently does first frames and doesn’t do the cinema prompting.
Next steps mentioned
- Grab BoxBox (especially CinePrompt) from the GitHub repo; possibly port BoxBox Studios to Electron and fork it.
- Finish this in a second session.
- Set up the project’s film style so camera choices follow it.
- Build a laptop element to control the screen shots.
- Drop renders into Premiere shot by shot and see what happens.
- Test the same prompts in other models and compare.
- The filmmaker to restore the character builder sheet Eric removed, and update the color tool.
Asides
- Floor mat. The filmmaker’s engineered hardwood floors were killed by a rolling chair after a year and a half. At Office Depot they found fabric floor mats, about $150: hard enough to roll on, soft enough not to be a sticky plastic mat. Rolling on this stuff usually sucks; this one rolls all day long. The visitor needs one too: they’re getting a room ready and will pull the rug back first.
- Scotch. Asked for water; joked about three, no, five fingers of scotch. “I used to be able to.”
- The prostate commercial. The filmmaker showed their last commercial project, made entirely with AI images (Midjourney for the concepts and finals, Kling for most video, one or two Midjourney videos; real men’s faces used). It was a huge hit. Lines from it: a former fighter pilot who thought he was strong and confident, facing a possible cancer diagnosis — “it’s like an acid to you, it burns you down and it wears you out.” Every year millions of men are told their PSA level is elevated; it’s an alarm that doesn’t say what it’s alarming for. Most are sent to an invasive surgical biopsy, and about 70–75% find no clinically significant cancer. One man’s first biopsy showed nothing; he spent six to eight months before getting an answer, including a saturation biopsy of about 50 needles, with high numbers the whole time. Another man’s father had prostate cancer, and it was ultimately the root cause of his death; his own cancer was lasered out within 24 hours, then 18 months later it recurred. The product: a blood test that gives physicians more information about clinically significant prostate cancer before the next step — “ask your doctor what your PSA really tells you.”
- Movie Maker Pro and a first film. A reel digitizer for 8 mm film. The filmmaker’s good friend, whom they met at 11, was over yesterday. As gifted and talented kids they spent a summer filmmaking at UT, and the friend’s mom saved the film — the filmmaker’s first film, on a reel in a box, to be scanned.
- Art camp. In fifth grade the visitor was sent to a gifted and talented summer art camp, the only one from their school. They drew sailboats in charcoal in Annapolis — hands dirty, hated it. They wanted to learn to drive tanks and jets instead. People often don’t have a perspective on that.
- HTML vs Electron. The visitor built an HTML server app for this, and for the third time found an Electron app would have been much easier. Electron just works, and pairs well with an MCP interface to Claude or omp. In omp they can render videos and infographics and use ten different models (ChatGPT, GLM, Grok…).
- Logistics. The recorder was briefly recording a window instead of the whole screen, and was fixed. A mic battery ran down fast and was put on USB-C to charge. Near the end someone checked the time with David; they were about done. “I like this tea a lot, by the way.”