Daniel sent us this one — he and Hannah are moving in two months, he's in the middle of decluttering, staring at rooms full of stuff, and his brain does what a lot of brains do when faced with chaos: it freezes. His idea is to use a walk-through video, downsample it intelligently to a handful of clean frames, and feed those to a multimodal AI that can actually tell him what to keep, what to toss, and how to group things. The question is whether tools exist that can pull that off — smart frame extraction that picks clean, non-blurry shots, ideally ones where something new has actually entered the frame.
By the way, today's script comes courtesy of DeepSeek V four Pro, so if anything sounds unusually coherent, that's why.
Or if it doesn't, we'll blame the model. But Daniel's question is genuinely practical. He's not asking for vaporware. He wants to know if there's something he can use right now, in May, to turn a three-minute phone video into maybe forty or fifty good frames and then get actionable decluttering advice out of a multimodal model.
The short answer is yes, the pieces all exist, but nobody has wrapped them into a single consumer app with a big friendly button that says "declutter my apartment." What Daniel's describing is a pipeline with three stages. Stage one: intelligent frame extraction. Stage two: loading those frames into a multimodal model. Stage three: prompting that model with the right system prompt. None of this is science fiction anymore. The interesting part is stage one — the smart downsampling — because that's where most people just throw up their hands and use something crude.
Daniel specifically called out the problem with naive frame extraction. If you just grab one frame every four seconds, you're going to get a lot of blurry junk. The frame where the camera is swinging between two points is useless. You want frames where the camera has stabilized on something — ideally where the content is substantially different from the last good frame you grabbed.
This is a solved problem in video engineering, just not one that's packaged for home decluttering. The tool that does exactly what Daniel wants — extract frames based on scene change detection, with a sharpness filter — is FFmpeg. It's a command-line tool, free, been around forever, and it has a specific filter called the scene change detector. You run it with a threshold, and it spits out keyframes where the visual content shifts enough to matter. Combine that with a blur detection filter and you've got exactly the pipeline Daniel described.
This is where Daniel's ADHD brain meets reality, because FFmpeg is not a friendly app. You type things like "ffmpeg -i walkthrough.mp4 -vf select equals gte scene comma zero point four, metadata equals print" and hope you didn't miss a flag. But there are wrapper tools now. I saw a write-up recently on Snowflake's developer guide about exactly this — extracting frames for multimodal AI analysis — and they walk through the FFmpeg pipeline step by step.
The Snowflake guide is a good reference point because it's aimed at developers doing video understanding at scale, but the principles are identical. You take a video, run scene change detection, output frames at a resolution multimodal models can handle — typically 512 by 512 or 1024 by 1024 — and feed those into Claude or GPT-4o or Gemini. The resolution matters because these models don't need 4K. They need enough pixels to recognize objects and spatial relationships, and beyond that you're just burning context window for no benefit.
Daniel mentioned this himself — he remembered our earlier discussion about how audio sent at too high a bitrate can actually degrade ASR accuracy because the model was trained on lower-resolution inputs. Video has a similar dynamic. If you send full-resolution frames, you're not just wasting tokens. You might actually get worse results because the model is trying to process detail it wasn't optimized for, and the important structural information — "this is a pile of books next to a lamp" — gets lost in the noise of individual pixels.
Token economics here are brutal. A single high-resolution image can be thousands of tokens. If Daniel shoots a three-minute walk-through at 30 frames per second, that's 5,400 frames. Even at one frame per second, that's 180 frames. If each frame costs 500 tokens, you're at 90,000 tokens before you've even written a prompt. Most models cap out well below that for image inputs, or they do aggressive internal downsampling you have no control over. So doing the frame selection yourself, before the model sees anything, is not just smart — it's necessary.
Let's get concrete. Daniel asked if there are programs on the market. The FFmpeg route works but it's nerdy. Are there more polished tools?
There's a company called Twelve Labs worth mentioning. They've built an entire platform around video understanding, and their whole approach is essentially what Daniel described — they don't process every frame. They use a "video-native" embedding model that extracts semantic information efficiently. It's aimed at enterprise use cases — searchable video libraries, content moderation — but the underlying technology is exactly the smart-downsampling-plus-AI pipeline Daniel wants. They have an API, so a developer could build a decluttering app on top of it pretty quickly.
There are consumer-adjacent tools too. I dug into this, and there was a piece on NewsBytes recently about AI tools specifically for home decluttering. The landscape is still fragmented. You've got apps that do object recognition on single photos — point your camera at a shirt and it tells you whether to keep it — but nothing that takes a walk-through video and gives you a room-by-room decluttering plan. The gap isn't the AI capability. The gap is the integration layer.
Which is fascinating because the AI capability is there now. If Daniel took his forty or fifty good frames and uploaded them to Claude or ChatGPT or Gemini with a prompt that says "I'm decluttering my apartment, here are frames from a walk-through video, identify items that appear to be clutter, suggest what to discard versus keep, and group items that should be stored together," the model would do a shockingly good job. I've tested things like this. These models can identify objects, assess whether a space looks cluttered, even estimate dimensions from reference objects. A standard door frame is about 80 inches tall — the model knows that, and can use it as a scale reference.
That dimension estimation point is one Daniel raised himself — he mentioned viewing a new apartment and estimating room sizes from a walk-through. And that works better than you'd expect. The model won't give you measurements down to the inch, but it can tell you "that living room looks about 15 by 20 feet based on the sofa and the doorway." For comparing apartments before you visit, that's useful.
Back to the frame extraction question, because I think that's the bottleneck Daniel is really asking about. There are a few paths he could take right now. Option one: use FFmpeg if he's comfortable with the command line. Option two: use a GUI tool like LosslessCut or HandBrake that can do scene detection and frame export, though they're not purpose-built for this. Option three: there are Python libraries — OpenCV has a scene change detection module, and there are pre-built scripts on GitHub that do exactly what Daniel described, picking the sharpest frame from each scene cluster. If he's an open-source developer, which he is, that's probably the sweet spot. Twenty lines of Python and he's got his frame extractor.
Daniel is an active open-source developer. He could absolutely write that script. The question is whether he wants to, in the middle of a move, with a toddler, while working full-time. Sometimes the right tool is the one that already exists.
And that's why I think the most practical answer for Daniel specifically is: use one of the existing multimodal chat interfaces, upload a video directly, and let the platform do the downsampling for you. Claude can accept video uploads now. It does internal frame extraction — not as controllable as doing it yourself, but for a three-minute walk-through, it's going to grab enough good frames to work with. The token cost will be higher than a hand-optimized pipeline, but the time savings might be worth it when you're two months from a move.
I think that's the trade-off Daniel is wrestling with. He's enough of a technologist to want the elegant solution — the perfect frame extractor that picks exactly the right moments. But he's also a guy with boxes to pack. The question isn't "can this be done" but "what's the path of least resistance that still works.
Let me give a concrete recommendation then, because Daniel asked for it directly. Step one: shoot the walk-through video on your phone, but shoot it with intention. Don't just wave the camera around. Pause for two or three seconds on each area you want the AI to see. Hold the phone steady. This alone will make any extraction method work better, because you're creating natural keyframes just by how you shoot.
That's a good tip that nobody thinks about. If you know the AI is going to sample frames, give it clean frames to sample. It's like speaking clearly for a voice assistant — you're adapting to the tool's constraints.
Step two: upload the video to Claude or ChatGPT or Gemini — whichever multimodal model you have access to. All of them handle video now. Step three: use a system prompt that gives the model a clear job. Something like: "You are a professional home organizer. I'm going to show you a walk-through of my apartment. For each room, identify items that appear to be clutter, suggest a sorting system — keep, donate, discard — and recommend how to group items for packing. If you see a pile of papers on a desk, tell me what to do with it.
The model will do that. It'll say "on the desk in what appears to be your home office, there's a stack of papers that looks like old bills and manuals — shred the bills, recycle the manuals unless they're for appliances you still own." That's the level of specificity these models can hit now. It's actionable.
The thing that makes this work where a human brain fails — and Daniel mentioned his ADHD here — is that the AI doesn't get overwhelmed. A human looks at a cluttered room and the visual input hits the brain as one giant undifferentiated "too much." The executive function required to break that down into individual items and decisions is enormous. The AI doesn't have that bottleneck. It just methodically catalogs what it sees and outputs a list. For someone with ADHD, that's not a gimmick. That's an accommodation that makes the difference between being stuck and being in motion.
We've talked about this before — the AI as an executive function prosthetic. It's not making decisions for you. It's doing the pre-processing that your brain struggles with, so you can make the decisions. "Here are the 37 items I see in this room, grouped by category, with a suggested action for each." Now your job isn't "deal with this overwhelming room." Your job is "look at this list and say yes or no to each line." That's a completely different cognitive load.
There's a company I've been watching in this space called Decluttr — not the electronics buyback service — that's been experimenting with AI-powered organization. But honestly, the general-purpose models have gotten so good that specialized apps are struggling to justify themselves. Why download a decluttering app trained on a specific dataset when Claude can look at your living room and tell you that the stack of mail on the coffee table should probably be sorted into "action required" and "recycling"?
Because the specialized app might have a better UI, might remember your preferences across sessions, might integrate with a to-do list or a moving checklist. The value isn't in the AI capability anymore — that's commoditized. The value is in the workflow integration. And that's the piece Daniel is correctly identifying as missing.
The pipeline exists. The individual tools exist. What doesn't exist — yet — is the single app that says "film your messy apartment and I'll tell you what to pack, what to toss, and how to label the boxes." Someone should build that. Moving is one of those universally stressful experiences where cognitive offloading has obvious value.
Daniel, being in the AI and automation space, is probably looking at this and thinking "I could build this." The question is whether he wants to build it during a move, or whether he wants to duct-tape together the existing pieces and get on with his life.
My vote is duct tape. Two months out from a move, with a kid, the last thing you need is a side project. Use Claude's video upload, get your decluttering plan, label your boxes, call the movers. Build the app after you've unpacked.
Let's talk about the specific frame extraction techniques Daniel mentioned, because there's a deeper technical point worth unpacking. He described two approaches. One is simple timed sampling — one frame every four seconds, with a sharpness filter to skip blurry ones. The other is content-aware — grab a frame whenever the scene has changed enough to be "new." These are both valid, but they serve different purposes.
The timed approach is simpler and more predictable. You know exactly how many frames you'll get — a three-minute video at one frame every four seconds gives you 45 frames. That's manageable for any multimodal model. The downside is you might miss things. If Daniel walks past a cluttered shelf in two seconds, a four-second sampling interval could skip it entirely.
The scene-change approach solves that — it grabs a frame whenever the visual content shifts, regardless of timing. But it has its own pitfall. If Daniel is walking through a room and the camera is panning slowly, the scene is changing continuously, and a naive scene-change detector might grab dozens of nearly identical frames, all slightly different but none adding new information.
Which is why the best approach is probably a hybrid. Use scene change detection to identify clusters of similar frames, then pick the sharpest frame from each cluster. That way you get one good representative shot of each distinct view, without duplicates or motion blur. OpenCV has a tutorial on exactly this — "scene boundary detection with keyframe extraction" — and the Python code is maybe 30 lines.
If Daniel doesn't want to write Python, there's another option. Some video players, like VLC, have a scene-snapshot feature. You can play the video, pause at each point you want to capture, and hit the snapshot button. It's manual, but for a three-minute video, you could do it in ten minutes. Sometimes the low-tech solution is the right one when you're under time pressure.
The manual approach also has a hidden advantage: Daniel knows his own apartment. He knows which views matter. An automated system might grab a great frame of an empty wall and miss the overflowing closet. Human-in-the-loop frame selection isn't elegant, but it's effective.
Let's shift to the second half of Daniel's question — the AI analysis itself. Once you have your frames, what do you actually ask the model to do? He mentioned two use cases: decluttering and apartment viewing. These have different requirements.
For decluttering, the model needs to identify individual objects and assess their state. "There's a pile of clothes on the chair — these look like they might be clean laundry that hasn't been put away versus dirty clothes that need washing." That's a judgment call requiring some reasoning. The model has to look at context clues — are the clothes folded, are they on a chair versus in a hamper, is there a laundry basket nearby.
The model can do that surprisingly well. I've seen demos where multimodal models distinguish between "this is a tidy stack of papers that needs filing" and "this is a scattered mess of receipts and junk mail that can be recycled." The difference is in the spatial arrangement, and these models are trained on enough images of messy versus organized spaces that they pick up the visual cues.
For apartment viewing, the use case is different. Daniel mentioned estimating dimensions. That's a spatial reasoning task. The model needs to identify reference objects of known size — doors, standard ceiling heights, floor tiles if visible, furniture of standard dimensions — and extrapolate. A standard interior door in the US is 80 inches by 30 inches. A kitchen counter is typically 36 inches high. If the model can see these reference points, it can estimate room dimensions with reasonable accuracy.
I'd add a caveat here. These estimates are ballpark at best. If Daniel is comparing two apartments and one looks significantly larger in the walk-through, the model can confirm that impression and give approximate numbers. But he shouldn't make a lease decision based on AI-estimated square footage. The error bars are too wide.
It's a screening tool, not a measurement tool. Use it to narrow down which apartments are worth visiting in person.
There's another use case Daniel hinted at that I think is underrated: the AI as a packing strategist. Once the model has seen all the rooms, you can ask it things like "what should I pack first?" or "which items across different rooms should be grouped together in the same box?" That's a planning task that benefits from seeing the whole apartment at once, which a human brain struggles to hold in working memory.
Oh, that's clever. The model sees the books in the living room, the books on the nightstand, the cookbooks in the kitchen, and it can say "you have books in three rooms — consolidate them into one box labeled 'books' rather than packing them separately by room." A human might not notice the pattern because the items are physically separated and encountered at different times during packing.
For someone with ADHD, that cross-room pattern recognition is exactly the kind of thing that's cognitively expensive. Your brain is focused on the room you're in. It's not naturally tracking "where else did I see books?" The AI does that automatically because it's processing all the frames as one dataset.
This connects to something broader about how multimodal AI changes productivity. We've spent years optimizing text-based workflows — to-do lists, notes, project management tools. But a lot of our cognitive load comes from the physical environment. A cluttered desk imposes a real tax on executive function. The ability to offload "what am I looking at and what should I do about it" to an AI is new.
It's the kind of thing that sounds trivial until you need it. If you don't struggle with clutter or executive function, the idea of asking an AI to look at your messy room and tell you what to do seems absurd — just clean it up. But for people who get stuck at the "where do I even start" stage, it's the difference between paralysis and action.
Daniel used the phrase "my ADHD speaking," and I think that's exactly right. He's not asking the AI to do the physical work. He's asking it to break the cognitive logjam. "Here's a list. Throw these things out, put these things here." That's scaffolding. It's the structure his brain needs to get into motion.
The beauty of it is that once he's in motion, he probably doesn't need the AI anymore. The hardest part is the transition from overwhelmed stasis to first action. The AI is a starter motor, not an engine.
Let me circle back to the technical question, because I want to make sure we actually answered what Daniel asked. He wanted to know if programs exist that do smart frame extraction for this purpose. The answer is: yes, but they're developer tools, not consumer apps. FFmpeg with scene detection, OpenCV with keyframe extraction, Twelve Labs' API for video understanding — these all exist and work well. What doesn't exist is the one-click "declutter my house from this video" app.
I think that's an honest answer. Daniel is enough of a technologist that he can probably work with the developer tools. But for listeners who aren't, the practical advice is: use the video upload feature in your preferred multimodal AI, shoot your walk-through deliberately with pauses, and craft a good system prompt. That gets you eighty percent of the way there with zero technical overhead.
The eighty-twenty rule applies here. The custom FFmpeg pipeline with hybrid scene detection and sharpness filtering is the hundred percent solution. The "upload video to Claude and ask nicely" is the eighty percent solution. For someone two months from a move, the eighty percent solution is the right one.
Unless you're the kind of person who finds joy in building the pipeline. And Daniel might be that person. He's an open-source developer. He might look at this and think "I could write a quick script, open-source it, and help other people in my situation." That's a valid choice too, as long as it doesn't add stress to an already stressful timeline.
If he does build it, I'd suggest keeping it simple. Python script, FFmpeg under the hood, outputs a folder of frames, done. Don't try to build a full app with a UI during a move. Ship the minimum viable pipeline and move on.
There's one more angle I want to explore. Daniel mentioned that his previous question about audio tokenization actually changed how he produces this podcast. He switched to sending transcribed text instead of raw audio because he learned that audio is token-heavy and the nuance wasn't worth the context cost. That's a real production change based on something we discussed. Now he's asking a similar question about video. It's a pattern — he's systematically figuring out how to use multimodal AI efficiently, not just as a consumer but as a builder.
That's why his questions are always interesting. He's not asking "is AI cool?" He's asking "how do I actually use this thing without burning through my context window and getting garbage results?" It's an engineering mindset applied to a consumer tool.
The video question is harder than the audio question, though. With audio, the optimization was straightforward — transcribe to text, send text, done. With video, the optimization is multi-dimensional. You're trading off frame rate, resolution, sharpness, scene coverage, and token cost. There's no single right answer. It depends on what you're trying to do.
That's why the system prompt matters so much. If Daniel is using the frames for decluttering, he might want higher resolution on cluttered surfaces and lower resolution on empty walls. A smart pipeline could do that — detect regions of interest and allocate resolution accordingly. But now we're talking about a research project, not a moving-day tool.
Let's bring this back to the practical. Daniel is moving in two months. He's got boxes, a labeling system, a wife and a toddler and a job. What should he actually do?
Here's my concrete recommendation, step by step. One: shoot a deliberate walk-through video on your phone. Pause at each area of clutter for three seconds. Hold the phone steady. Two: upload the video to Claude. Three: use a system prompt that says something like "I'm moving in two months and need to declutter. For each room in this walk-through, identify items that can be discarded, donated, or need special packing. Group items that should be packed together. Be specific about what you see and what action to take." Four: take the output and use it as a working checklist. Don't treat it as gospel — override anything that doesn't make sense. But use it as a starting point so you're not starting from a blank page.
If the video upload burns too many tokens or hits a limit, fall back to the manual approach. Take 10 or 15 well-framed photos with your phone — one per area of interest — and upload those instead. Photos give you more control over what the model sees and cost fewer tokens than video. For a three-room apartment, 15 photos might be all you need.
That's actually the simpler approach and I should have led with it. Photos are easier to control, cheaper in tokens, and give you the same analytical capability. The only thing video adds is the walk-through flow, which is nice for capturing spatial layout but not essential for decluttering. For apartment viewing, the walk-through matters more because spatial flow is part of what you're evaluating.
For the apartment viewing use case, Daniel might actually want to keep the video format, because "how does it feel to walk from the kitchen to the living room" is information that still photos don't capture. The model can pick up on things like narrow hallways, awkward transitions, whether the rooms flow well together.
That's a good distinction. Decluttering: photos are probably better. Apartment evaluation: video walk-through has advantages. Different tools for different jobs.
One more thought on the apartment viewing side. Daniel could use the same pipeline to compare multiple apartments. Shoot walk-throughs of three different places, extract frames from each, and ask the model to compare them across dimensions — natural light, storage space, room sizes, layout efficiency. That's a comparison task that's hard for humans because you're relying on memory of each visit. The AI holds all the visual data simultaneously.
That's actually brilliant. You visit three apartments in a day, they blur together in your memory, and by the evening you can't remember which one had the bigger kitchen. Feed the walk-throughs to a multimodal model and ask for a structured comparison. It'll tell you "apartment A has the largest kitchen but the least closet space, apartment B has the best natural light, apartment C is the most efficient layout." That's actionable.
It's the kind of thing that sounds like a gimmick until you've apartment-hunted in a tight market where you're seeing five places in a weekend. The cognitive load of tracking all those details is real.
We should probably mention the privacy consideration here, because Daniel is uploading video of his home — and potentially homes he's viewing — to a cloud AI service. For his own apartment, the clutter and personal items are visible. For apartments he's viewing, there might be the current tenant's belongings in the frame. That's worth being thoughtful about.
For his own place, he's probably comfortable with it — he's already sending audio of himself and his family to AI services for the podcast. For apartments he's viewing, he should be careful not to capture anything that would violate someone else's privacy. Wide shots of empty rooms are fine. Close-ups of someone's family photos on the fridge are not.
Practically speaking, if he's using a cloud service, the video is being processed on someone else's servers. The major providers have privacy policies that say they don't train on user uploads, but if there's anything sensitive — financial documents on a desk, medication bottles — he might want to clear those out of frame before shooting.
Or use a local model. There are open-source multimodal models now that can run on a reasonably powerful laptop. 2 has vision capabilities. You can run it locally with something like Ollama. It won't be as good as Claude or GPT-4o for this task, but it keeps everything on-device.
The local model option is interesting but I'd caution that the quality gap is still significant for detailed visual analysis. A local model might correctly identify "a pile of clothes" but miss the distinction between "clean folded laundry" and "clothes that need washing." For decluttering, that distinction matters. For now, the cloud models are substantially better at this kind of fine-grained reasoning.
The trade-off is privacy versus capability. Daniel has to make that call based on what's in his video. If it's just boxes and furniture, the privacy concern is minimal. If there's personal paperwork visible, maybe do a quick tidy before shooting.
Alright, let's zoom out and talk about what this means for the broader landscape. Daniel's question is specific — "does this tool exist?" — but it points to something bigger. We're in this weird transitional moment where the AI capabilities are clearly there, but the integration layer is missing. The pieces are on the workbench but nobody's assembled them into a product.
That's frustrating if you're a consumer who just wants the product. But it's exciting if you're a developer who sees the gap and knows how to fill it. Daniel is both — he's a consumer with an immediate need, and a developer who can see the product that should exist. The tension in his question is between "I need this now" and "I could build this.
My advice to developer-Daniel: build it after the move. My advice to consumer-Daniel: use the duct-tape solution now. Claude video upload, deliberate walk-through, good system prompt. It'll work well enough to get the boxes packed.
If he wants to get slightly fancier without writing code, there are intermediate options. Some photo management tools have AI-powered tagging now — Apple Photos, Google Photos — and they can identify objects in images. It's not a decluttering plan, but it can help you catalog what's in each room. "Search for 'books' and see all the photos that contain books across different rooms." That's a manual version of the cross-room pattern recognition we talked about.
The Apple Photos approach is actually pretty good for this. The object recognition is on-device, so privacy is preserved. You can search for "boxes," "clothes," "papers," "cables" — all the clutter categories — and it'll surface photos containing those items. It won't tell you what to do with them, but it'll help you see patterns you might have missed.
For the apartment viewing use case, there's an app called MagicPlan that uses your phone's camera and AR to create floor plans with measurements. It's not AI in the LLM sense, but it solves the dimension estimation problem more accurately than a multimodal model would. Daniel could use MagicPlan for measurements and Claude for qualitative assessment — "how does this layout feel?" — and get the best of both worlds.
MagicPlan is a good call. It uses the phone's LiDAR sensor if you have one, or just the camera with clever computer vision. The measurements are surprisingly accurate — within a few inches in most cases. For comparing apartments, having actual floor plans with dimensions is way more useful than AI-estimated guesses.
The toolkit Daniel might actually assemble is: phone camera for deliberate photos or walk-through video, Claude or similar for the decluttering analysis and apartment comparison, MagicPlan for floor plans and measurements, and maybe Apple Photos for on-device object search across rooms. None of this requires writing code. It's all off-the-shelf.
If he does want to write the frame extraction script later, he can. But the off-the-shelf toolkit covers his immediate needs for the move. That's the pragmatic answer.
I think we've given Daniel a pretty thorough answer. The specific tool he asked about — a smart frame extractor that picks clean, novel frames from video — exists in the form of FFmpeg and OpenCV scripts, but not as a consumer app. The practical path for someone in the middle of a move is to use existing multimodal AI tools with deliberate input — good video or photos, good prompts — and get the decluttering or apartment-analysis output he needs.
The deeper point, which I think is what makes this worth discussing beyond just Daniel's situation, is that we're at a moment where AI can serve as an executive function prosthetic for tasks that overwhelm human cognition. Decluttering, comparing apartments, planning a move — these are cognitively heavy tasks that benefit enormously from having an external system break them down into manageable steps. The technology exists. The integration is lagging. But even without perfect integration, the pieces are good enough to be useful today.
Now: Hilbert's daily fun fact.
Hilbert: The scoring system in real tennis, the medieval precursor to lawn tennis, counts by fifteens — fifteen, thirty, forty-five — and the term "deuce" likely derives from the French "à deux," meaning the game is two points from completion. The name "tennis" itself probably comes from the French "tenez," meaning "hold" or "take this," which players would shout before serving. This sport was played in enclosed courts across medieval France and England, with complex rules involving angled walls and penthouses, and the current world champion as of twenty twenty-six is Camden Riviere.
...right.
For Daniel and anyone else staring at boxes and clutter: the AI can't pack for you, but it can tell you where to start. And sometimes that's enough. This has been My Weird Prompts. Thanks to our producer Hilbert Flumingtop. If you want more episodes like this one, head to myweirdprompts.We'll be back soon.
