초보와 고수의 프롬프트 차이 | 클로드 Fable 5.2 화제작 50개 — Transcript
Full transcript
- 0:00On November 9, Fable 5.2 was released, and the next day Google launched Gemini 4 Carbon.
- 0:07Over the next six days, X was flooded with works made with the two models.
- 0:11This video presents the 50 most popular ones, ranked by likes.
- 0:16We read each post, the creator's replies, and shared prompts and repos, then summarized how each was made and how to replicate it.
- 0:25Posts that used other tools or include promotion are marked on screen.
- 0:30Let's start with #50.
- 0:36A calligraphy toy: write a character and ink spreads across the screen.
- 0:40The ink is simulated as a real-time fluid.
- 0:43It was an open request: no set subject, just "make something that shows off your abilities."
- 0:49But commenters noted that the same request gave very different results.
- 0:55A post from the CAD company MecAgent.
- 0:58Feed in a 3D design model, and Fable 5.2 draws the 2D drawings.
- 1:03It ran on the company's in-house agent that drives CAD software, and some countered that CAD already had automatic drawings.
- 1:14A voxel engine built from a single prompt, still playable on the web today.
- 1:19The model reportedly ran its own loop: improve quality, optimize, then improve quality again.
- 1:25The creator offers two tips.
- 1:27To build something big in one prompt, the prompt must be perfect, and low or medium reasoning effort is usually enough.
- 1:37The process has two steps.
- 1:39First, an image AI drew one screen of an alien-planet shooter. Then Claude Code got the image: "Make a three.js game from this."
- 1:50The colors and layout are all in the image, so a one-line prompt is enough.
- 1:55Reasoning effort: medium. About $13 at API rates.
- 1:59Some comments questioned the 30-minute time.
- 2:04The same robot arm was connected to Fable 5.2 and GPT-6 Astra, and each painted The Starry Night from the same prompt.
- 2:13Human help, like reinserting brushes and refilling paint, was disclosed, and the verdict: Astra still leads at robot tasks.
- 2:23A shopping site where you flip through clothes on a rack.
- 2:27It started from a one-line concept: a rack instead of a grid. The prompt is on the creator's paid site.
- 2:35Practitioners noted it's unlikely to drive purchases, and on phones only a few garments are visible.
- 2:43Here, Fable 5.2 is the director.
- 2:46It wrote the script, read the storyboard, and planned the workflow, while the video
- 2:50was generated by Seedance 2.5 in Filmera, connected via MCP.
- 2:55Asked why the motion looked robotic, the creator said to write out each character's acting
- 3:00and tell them to move dynamically, like Pixar or Disney.
- 3:04The secret to keeping faces consistent is a character sheet.
- 3:09A photorealistic bedroom made with three.js code alone.
- 3:13Instead of real-time lights, skylight from the window was pre-baked with path tracing, so it looks real even on laptop graphics.
- 3:22The published rules say to fix the random seed and define every dimension in meters.
- 3:30First, the main agent set the streets, building heights, sun position and camera
- 3:35in a single layout file, and blocked out the scene with gray boxes.
- 3:39Then ten subagents split up the work: skyscrapers, vehicles, rooftop props.
- 3:45The rule: each edits only its own file and pulls numbers from the layout file.
- 3:51That's why the ten parts reportedly fit together on the first try.
- 3:54Windows were offset by a few millimeters, and 5% of roof tiles were removed on purpose.
- 4:00What the company recommends trying first is the lighting.
- 4:04One low sun and a light haze, they said, do more than adding detail.
- 4:11Da Vinci's flying machine was built in 3D, stress-tested under physics, then redesigned to fix its weak points.
- 4:19Build, test, fix: that's the order.
- 4:23But the physics engine and test results weren't shared, and the post ends with an ad for a free trial.
- 4:31From the first ten, there are three templates you can use right away.
- 4:35One: image first.
- 4:37Make the screen you want with an image AI, then ask it to build exactly that.
- 4:42Two: a layout file.
- 4:44Set dimensions and positions in one file first, and let each subagent edit only its own file.
- 4:50Three: acting directions.
- 4:52Write out how each character should move.
- 5:00Ask Claude for music, and it opens this composing app and layers tracks.
- 5:05Fable 5.2 built the app itself and released it as open source.
- 5:10The trick is how it connects.
- 5:12No extra connector: the song file is JSON the AI reads and writes, so editing the file alone creates the song.
- 5:20Features were verified by measurement, not by ear.
- 5:25One HTML file, 97 KB, holds a kelp forest, schools of fish, and a stone whale-tail sculpture.
- 5:34The shared prompt lists priorities by number.
- 5:38First visual quality, second frame stability, third file size.
- 5:42And an end condition: fix until a critic subagent that sees only the screen, not the code, says only minor issues remain.
- 5:52A jet assembles from 931 parts and 24,260 rivets.
- 5:58The company's first tip: lock down the scope first.
- 6:02Changing the length and part count midway cost them an hour.
- 6:06Three final-quality stills were approved before animation, and since the agent can't see motion, ffmpeg measured flicker.
- 6:18The shared prompt is just a few lines; two sentences are the core.
- 6:23"Send me screenshots while you build."
- 6:25"Use subagents to critique it."
- 6:28Giving each room its own subagent was the model's own decision.
- 6:32It used up the daily limit and continued the next day, about $315 at API rates.
- 6:41A game made in Unreal 5.9, with Claude Code driving the engine directly.
- 6:46Concept art by image generation, characters by Tripo, props in Blender, sound by ElevenLabs.
- 6:53Notably, all the logic was written in C++, not Blueprints.
- 6:59A choice made so the AI can read and compare the code.
- 7:04Time taken: 10 minutes. Cost: $8.43.
- 7:08But the shared prompt is a spec of nearly 6,000 characters.
- 7:12Only WebGL2 and Web Audio, no libraries; even the glass knob's magnification and the sound frequencies were given as numbers.
- 7:21The backbone: the knob's position is both the time of day and the story.
- 7:26Push it further and dawn turns to morning, as the music builds layer by layer.
- 7:31A wuxia action game made in about 6 hours with 450 million tokens.
- 7:37It wasn't one-shot.
- 7:39The creator played it seven times, flagging problems each time, and shared that experience as a skill.
- 7:45The skill's first sentence is the key.
- 7:48Projects like this fail not because of the code, but because no one looks at the screen, listens to the sound, or presses the keys.
- 7:58Glowing booster effects added to a 3D avatar.
- 8:02The character was made with the 3D AI Tripo, and the agent ran Blender via MCP, so the creator never opened it once.
- 8:11Values like effect length were exposed as sliders for a human to tune.
- 8:15When the model tried to build everything itself, the creator set a ratio: 90% must use Tripo models.
- 8:22Babylon in 600 BC, built over 8 hours.
- 8:27Fable reportedly looked up history texts and building plans itself, and fixed the scene by comparing it with them.
- 8:34But some objected that the same buildings were laid out in a grid, and the sources used weren't disclosed.
- 8:41A paper diary brought into 3D on the web.
- 8:44Usage was limited, so the work was split between models.
- 8:48Opus 5.5 got all the features working first; then Fable 5.2 polished only the materials against video of a real diary.
- 8:58In effect, making it work and making it look real were solved separately.
- 9:03Three templates from this section.
- 9:06One: add a critic.
- 9:08Have it take screenshots while building, and let a screen-only subagent critique them.
- 9:13Done when only minor issues remain.
- 9:16Two: priorities and numbers.
- 9:18Rank quality, frame rate, and file size by priority, and write performance targets as numbers.
- 9:25Three: cheap model first.
- 9:27Finish the features with a cheaper model, and use the expensive one only for final polish against the real thing.
- 9:37The creator cited $37 in tokens and a 30-minute render.
- 9:42Yet the shared prompt says no enemies or violence, while the video shows shooting.
- 9:49Many comments said it was real game footage, and there's no link or code.
- 9:54Comparing the prompt with the result is also a way to read a work.
- 10:00Show your face, and the skull and muscles move along.
- 10:04It's not a real X-ray: open anatomical 3D data is fitted to your head in real time by face-tracking AI.
- 10:11Combine open science data with tracking AI, and just describing it in words yields a tool like this.
- 10:18Your video never leaves the browser.
- 10:22An auto-battler where a lantern-bearing party crosses the dark; every model and animation is code.
- 10:29The creator's tip is one line.
- 10:31Use skeletal animation, analyze the walk frame by frame, and let it fix itself until it judges the walk correct.
- 10:38In about 5.5 hours, the creator stepped in 16 times to redirect, about 45 minutes total.
- 10:46Fable 5.2 directed, and Opus 5.5 subagents did the work.
- 10:52The public repo shows each subagent handles one file, and shared files are marked "do not edit."
- 10:59Every scene is a function of time, so only fixed frames need re-rendering.
- 11:04Some comments questioned the 30-minute time.
- 11:091.04 million nanobots morph into a hand, a whale, and a rocket.
- 11:14The order is the trick.
- 11:16Design one bot first, then the room, and finally the swarm.
- 11:20Neighbor checks look at one representative per cell to cut computation, which is how it runs at 60 fps.
- 11:28A two-legged robot that walks by computation, with no pre-made animation.
- 11:32It ran hundreds of simulations to find and fix where it fell.
- 11:37Walking and running were finished first in a light three.js prototype, then ported to Unreal 5.8 once it looked right.
- 11:45Validate in a prototype, then move to the engine.
- 11:50A documentary-style clip of dinosaurs grazing by the water.
- 11:54Fable 5.2 drove Blender with Python code, reportedly with no image or video generation models.
- 12:01Almost every Blender feature can be called from Python.
- 12:05The dinosaur models' source and the prompt weren't disclosed.
- 12:10A browser game modeled on the real Earth, with weather synced to reality in real time.
- 12:16The same creator as #48 shares more tips.
- 12:20Let one agent see it through rather than splitting agents, and lock in the art style first by showing lots of examples.
- 12:27The opposite of #42's ten-way split.
- 12:31Many comments also reported it failing to load.
- 12:35A motion graphics video telling editors, "I'm coming for your job."
- 12:4020 minutes, one request: the creator called it 100% Fable, music included.
- 12:46But the creator admitted that this model can't hear sound,
- 12:49and many comments also criticized the audio.
- 12:54Every single flower is 3D.
- 12:57From the maker of Mixar, a Blender-based 3D editor: Fable 5.2 built the scene inside it and rendered it with Cycles.
- 13:07The method was described only as "recursive prompting," and the team's artist handled the direction.
- 13:14Three templates from this section.
- 13:17One: motion self-check.
- 13:19Use skeletal animation, analyze the walk frame by frame, and fix until it's judged correct.
- 13:26Two: start small.
- 13:28Grow from one part to a small group to the whole, and validate in a light prototype before moving to the engine.
- 13:35Three: set the style with examples.
- 13:38Lock in the art style first with several example images, not words.
- 13:47An animation that redraws the same ocean scene at the graphics level of each year.
- 13:52Given only a one-line theme, the creator said Fable did 99% of the creative work.
- 13:58But many pointed out it differs from real history, such as jumping straight from 2001 to 2026.
- 14:05For fact-based topics, it's safer to provide a timeline first.
- 14:11The only Gemini work in the top 50.
- 14:14It actually runs Windows 11 ARM on a foldable phone.
- 14:18In 2 hours, an agent swarm built not an app but an entire virtual PC, including an instruction translator and virtual devices.
- 14:27The code is public, but to replicate it, you need an Apple silicon Mac and a Windows installer.
- 14:34A 90-second film of red and white dragons in flight, a single HTML file with no image, model, or audio files.
- 14:42Three steps.
- 14:43Give a Midjourney image as reference, have it recreated exactly in code, then iterate for 12 hours, saying only what looks off.
- 14:51Agents saved tokens by reusing each other's work via a shared bulletin board.
- 14:58A video introducing the creator's own service.
- 15:00Only Fable 5.2, no skills or MCP; the music was reportedly chosen in advance.
- 15:06Asked for the prompt, they only replied that it was a process, not a prompt.
- 15:11The 10 minutes is run time; the preparation wasn't disclosed.
- 15:17A game where you sail a boat through eight dreamlike worlds.
- 15:21Fable 5.2 scored the trailer using royalty-free instrument samples, while Suno made the game music and Tripo the characters.
- 15:30With recorded samples, even a model that can't hear composes with real instrument tones.
- 15:37It starts from a hand-drawn sketch.
- 15:40World Labs Atlas built the space as 3D Gaussian splats; Fable 5.2 used three.js for first-person walking, style shifts, and post-processing.
- 15:52Another tool built the space; Fable built the experience.
- 15:57A 35-second video of PC parts assembling: 7 hours 41 minutes, $138.82.
- 16:05The creator disclosed the full method.
- 16:08The main agent only manages, and a single subagent makes the video.
- 16:12The plan is approved first, then the target image and test renders
- 16:16are compared through at least three revision rounds before the final render.
- 16:20A sequel split work to Opus 5.5, but costs only fell to $118.
- 16:28A free site for learning eight real boards by rotating them in 3D.
- 16:33Every board is drawn in code, with no photos or 3D files, down to the traces and silkscreen text.
- 16:41Missions include placing parts, soldering, and wiring circuits, but critics said mistakes like cold joints are hard to learn in a sim.
- 16:51A 4-minute music video from the request "make the video you dream of," with its production log made public.
- 16:58Sets were SVG in code, motion by Seedance, songs by Suno; the human only asked, approved the plan, left one note, and picked the song.
- 17:10The first version's failure was the song.
- 17:12Fable, unable to hear, picked the worst track from descriptions alone, so a rule emerged: humans pick songs by ear.
- 17:22A one-line request in Japanese.
- 17:24"Show me the ultimate video that shows what you can do now."
- 17:28Reasoning effort was medium, and unprompted, Fable started by composing an original song.
- 17:34The creator offers two tips.
- 17:36Fable is slow, so hand it everything you want at once, and split the work: Fable designs, Opus or Codex implements.
- 17:46Three templates from this section.
- 17:48One: manager-style direction.
- 17:51The main agent builds nothing itself; it hands subagents a complete brief and only inspects the results.
- 17:58Two: recreate the image in code.
- 18:00Have it match a reference image exactly, and when revising, just point out what looks off.
- 18:06Three: humans pick the music.
- 18:08The model can't hear, so a human picks the song by ear and the model matches it.
- 18:19A video depicting Claude working through the night.
- 18:21The trick is the laptop screen.
- 18:24Not a recording: a web page that draws any given moment was captured as 1,345 frames and put on the screen in Blender.
- 18:32One file of typing timestamps was read by the screen, keyboard, and sound effects together to keep them in sync.
- 18:39Rendering was spread over 10 GPUs for $5.75.
- 18:43However, the Claude usage cost wasn't disclosed.
- 18:47It turns a floor plan into a 3D house you can walk through.
- 18:51The shared prompt's first principle: the builder never grades its own work.
- 18:56Each room's render and real photos go to a fresh critic that never saw the work, returning numeric fixes until the room passes.
- 19:06But many comments doubted the dimensional accuracy, saying a wall off by just 30 cm gives it away.
- 19:14A browser open world roaming part of Manhattan.
- 19:18It was billed as "Rockstar: 8 years, Fable: 7 days," but even the creator said it's not GTA.
- 19:25The key instruction was one line: use a gauntlet loop.
- 19:29A blindfolded AI compares the scene with a real NYC photo and picks the real one; what gave it away is fixed next round.
- 19:38After launch, it took player feedback and shipped five updates in a day.
- 19:44An agent team ran a Meta ad campaign from scratch through publishing.
- 19:49It drove the browser directly instead of the API, and the creator recommended a daily
- 19:54budget cap, with anything above it needing human approval.
- 19:58But doubts about authenticity, such as the account being a demo, went unanswered.
- 20:06A game where you walk through a forest that looks painted in oil.
- 20:09Not a filter: one brushstroke is placed at each point in 3D space.
- 20:14About 1.5 million per frame.
- 20:16Two tips.
- 20:18Strokes stay fixed in the world so they don't slide as you move, and a jitter and frame-time test script lets the model check in numbers.
- 20:28200 fps is on an RTX 4090.
- 20:3424 million views, the most-watched work.
- 20:38Halving the distance step by step, it zooms to atomic scale to show whether a finger truly touches.
- 20:44The atoms are a real keratin structure from the Protein Data Bank, and the ruler is to real scale.
- 20:50Making it use real data was the trick.
- 20:52But a Community Note said it mixed a paradox with atomic physics, and the creator agreed.
- 21:00A shooter made by an Anthropic employee.
- 21:03The request was one paragraph.
- 21:05"AAA quality, but keep the scope to one small level."
- 21:09Then 8 hours with no human feedback, about $550 at API prices, with most input served from cache.
- 21:17Textures and models are all shader code, so the whole game is one 1.4 MB HTML file.
- 21:24But it continued an earlier conversation, sparking debate over whether it was one-shot.
- 21:30Anthropic's official launch post, made by Kevin Ngo of #16.
- 21:35Instead of asking for the cell all at once, it built up organelles one by one.
- 21:39Note the explanation cards.
- 21:41They call out where the drawing departs from reality, like "actually 5 nm thick, but drawn much thicker."
- 21:49That's how to keep scientific visualization accurate.
- 21:53A game where Van Gogh rides a bike through The Starry Night.
- 21:58The code shows the sky's center is a scan of the painting; beyond it, shaders extend it with the original's measured colors and stroke angles.
- 22:07It extends the original instead of discarding it.
- 22:10Code comments record abandoned approaches along with why they were dropped.
- 22:15Comment requests became features within a day, but the backlash against AI art was fierce.
- 22:22#1 is something that happened off screen.
- 22:24A 1994 Jeep with no diagnostic port was hooked to a $3 Arduino; Claude Code handled diagnostics and Codex the memory work.
- 22:34Humans only did the wiring and driving; Fable 5.2 was used to edit this video.
- 22:40The know-how is in the public repo.
- 22:42It includes instructions for agents: reading is safe, but three other actions require asking every time.
- 22:51Templates from the final ten.
- 22:53One: a fresh critic.
- 22:56So the builder doesn't grade its own work, get numeric fixes from a critic seeing it for the first time.
- 23:02Two: spot the real one.
- 23:05Show the built scene and a real photo side by side, have it pick the real one, and fix what gave it away.
- 23:12Three: ask before risky actions.
- 23:14Let it read freely, but a human approves anything irreversible or anything over a set amount.
- 23:22Of the 50, 48 were made with Fable 5.2, two of which also used Opus 5.5.
- 23:31One was a Gemini 4 Carbon work, and one was a video comparing two models.
- 23:3621 shared a link or code you can try, and 15 also used other AI tools.
- 23:44Stated build times ranged from 10 minutes to two days, and costs from $8.43 to about $550.
- 23:54We tagged and counted the techniques confirmed in each work.
- 23:58#1 is upfront planning, used in 19 works.
- 24:02Dimensions, order, and priorities were set before building.
- 24:06#2 is combining other AI tools, and #3 is a visual feedback loop that checks results on screen and fixes them.
- 24:13Nine were one-shots, with requests ranging from one line to a 6,000-character spec.
- 24:22Let's read four shared prompts, color-coded by role.
- 24:27#47 is two lines.
- 24:29Blue is reference, orange is the goal, purple is the tech choice.
- 24:34The image explains the colors and layout, so the text only needs to say what to build and with what tech.
- 24:43In #37, two sentences did the work.
- 24:46The first is goal and reference: one famous game map's name conveyed the era and mood.
- 24:52The second, in green, is verification.
- 24:55"Send screenshots while you build, and have a subagent critique them."
- 25:01#39 is in spec format.
- 25:04Yellow is priority: in a conflict, choose quality, then frame rate, then file size.
- 25:10Orange is numeric targets: a set device and resolution, and an average of 11 ms or less.
- 25:16Green is verification, lime is the end condition.
- 25:20With an end condition, the model knows where to stop.
- 25:25#35 is a spec of nearly 6,000 characters.
- 25:28Even materials and sound are in numbers; the part in pink is the backbone.
- 25:33The knob's position is the time of day and the stage of the story.
- 25:37Behind the result of 10 minutes and $8.43 was this much design.
- 25:44One more video from outside the 50.
- 25:47Same Fable 5.2: the top is the final prompt, the bottom is the first prompt's result.
- 25:53Though the creator said the first prompt was written sloppily on purpose.
- 25:58What to watch is the order of fixes.
- 26:00From the first result, three follow-ups came in order: structure, detail, then light and camera.
- 26:07The final prompt spelled out identifying features in numbers, like the seven rays of the crown.
- 26:15Reading all 50 also revealed limits.
- 26:18First, the same request gives different results, so it's hard to reproduce.
- 26:23Second, hidden costs and time: some cite only render costs or omit the prep.
- 26:29Third, claims and evidence sometimes differ, like a video that contradicts its shared prompt.
- 26:35Fourth, the models can't hear sound.
- 26:38Fifth, factual accuracy, such as years and dimensions.
- 26:42Sixth, other tools made much of the work, so you must separate out what the model did.
- 26:50One more, as a bonus.
- 26:52A pelican on a bicycle passes 20 frames, changing art style with each one.
- 26:57But the original creator is unconfirmed, and claims about the model are split between Gemini 4 Carbon and Opus 5.5.
- 27:06The shared prompt is worth studying.
- 27:08Build shared tools and specs first, write an art brief per panel for subagents, and stop at each stage to show a demo.
- 27:20After reading all 50, what the best works had in common was clear.
- 27:24Design first, give references, and let it check and fix its own results.
- 27:30And set an end condition and checkpoints for human review.
- 27:34Today's prompt templates and links to all 50 original posts are in the description.
- 27:41If this helped, please subscribe and like.
About this transcript
This page contains the full transcript of 초보와 고수의 프롬프트 차이 | 클로드 Fable 5.2 화제작 50개 by 사장님은 말만 합니다, generated from the public captions YouTube serves with the video. The transcript has 3,804 words across 306 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.