15:57Mansel Scheffel · Tutorial
Claude Cowork Projects Just Changed Everything (AI OS)
A 16-minute screen-share tour of how to build a four-department AI operating system inside Claude Cowork Projects — no IDE required.
March 22ndThe argument in one line.
Claude Fable sub-agent architecture isolates each tool task in its own context box, letting a fully autonomous session build complex software for hours without hitting limits — and prompt caching cuts the apparent token bill by 90%.
Read if. Skip if.
READ IF YOU ARE…
- You are a builder or developer curious about how far a single autonomous Claude session can go on a real software project.
- You want a concrete cost breakdown — tokens, cache rates, PixelLab credits — before committing to a Fable-scale experiment.
- You are a non-expert in a domain (game dev, design) wondering whether a detailed spec file is enough to delegate a full build.
- You are trying to understand sub-agent architecture and why it prevents context blowup in long autonomous runs.
SKIP IF…
- You want a step-by-step tutorial on how to replicate this — the video is observational, not instructional.
- You are looking for game development fundamentals — this is an AI capabilities demo, not a game-dev primer.
The full version, fast.
A non-game-developer gives Claude Fable a spec file, a PixelLab MCP connection, and one prompt, then lets it run for three hours unsupervised. The result is a working browser-based Stardew Valley clone with farming, fishing, mining, NPCs with contextual dialogue, day/night cycle, weather, and a farmhouse interior — built across three prompt rounds for $316.24 at API rates. The economics work because 99.1% of the 258 billion tokens consumed were cache reads at one-tenth the input price, and sub-agent isolation kept the main context from ever hitting its limit.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:44
01 · Cold open and project framing
Creator declares genuine amazement, establishes premise: build Stardew Valley in the browser using Claude Fable and PixelLab MCP. Not a game developer.
01:44 – 03:57
02 · Model setup and why PixelLab
Explains model choice (Fable, extra high), PixelLab MCP rationale, the Sproutlands asset pack alternative, and the $24 PixelLab tier. Launches the build.
03:57 – 08:19
03 · Spec file walkthrough
Reviews the spec: 10 done criteria, tech contract (browser-based after Opus recommended it three times), controls, world systems, PixelLab integration guide, and build order.
08:19 – 11:14
04 · Fable building — asset phase
Fable generating sprites in parallel via PixelLab. Creator discusses best use cases for Fable while it runs.
11:14 – 15:58
05 · Mid-build and sub-agent explanation
Still building at 35 min. Fable self-corrects failed tilesets. Creator explains how sub-agents prevent context blowup.
15:58 – 19:00
06 · v1 gameplay demo
First live demo of Claude Valley — chopping trees, breaking rocks, catching fish. Sound effects were unrequested.
19:00 – 22:15
07 · Token and cost breakdown
Deep dive: 731M tokens, $162.94, 99% cache reads. Explains why $163 not $2,000. PixelLab: 156/5,000 credits = $1.40.
22:15 – 26:06
08 · v2 build — polish and world dressing
Second prompt: audio, FX, pro farmer sprites, NPC animations, daemon task board with stall detection. Creator feeds Stardew screenshot for color reference.
26:06 – 30:00
09 · v2 gameplay, final tally, philosophy
Demo: rain, NPC dialogue, shop, cave mining with infinite levels, a cat. Final cost $316.24 across three versions. Closes on AI, craft, and human expression.
Lines worth screenshotting.
- 99.1% of Claude Fable tokens in a 3-hour autonomous build were cache reads billed at one-tenth the input rate, dropping a theoretical $2,000 bill to $316.
- Sub-agents quarantine context — each tool task runs in its own isolated window — which is why a 3-hour session never once triggered compaction.
- A non-game-developer with a spec file and an MCP-connected asset generator can get a playable prototype faster than hiring a junior dev to scope the project.
- Claude added a cat in the farmhouse, contextual music changing when fishing, and NPC dialogue referencing the player stated goal — none of this was in the spec.
- PixelLab $24/month flat tier generated the entire game sprite library for $1.40 worth of API credits.
- Setting a model to extra high compute instead of max with workflows delivers better cost-to-result ratios on long agentic tasks.
- A daemon task board with stall detection and heartbeat telemetry cut the second build phase time significantly — the model built its own watchdog.
- The cache hit rate scales with session length: the longer the autonomous run, the higher the proportion of cheap cache reads versus expensive fresh input.
- When AI handles the domain expertise you lack, human value shifts to spec quality, integration choices, and the irreplaceable expression layered on top.
- Binary done criteria written into a spec file before launch give the model a self-verification target and reduce mid-session drift without interruption.
Cache rate is the number that actually matters.
WHAT TO LEARN
A quarter-billion tokens sounds catastrophic until you see that 99% were cache reads billed at one-tenth the price — the hidden variable that makes long autonomous agent sessions economically viable.
- Cache reads cost roughly one-tenth of fresh input tokens; a session with a 99% cache hit rate bills like a session one-tenth its apparent size.
- Sub-agent isolation enables hours-long autonomous runs without context compaction — each tool task lives in its own context box and only returns its result.
- Writing binary done criteria into a spec file before launching an autonomous agent gives the model a self-verification target and reduces mid-session drift.
- When a model is going to run unsupervised for hours, ask it to build its own stall-detection daemon — the creator round two finished significantly faster after adding heartbeat telemetry.
- The ceiling on what you can delegate is set by spec quality, not domain expertise — the creator was not a game developer and got a playable game with fishing, mining, NPCs, and weather.
Terms worth knowing.
- Sub-agent
- A child agent instance spawned to handle a specific task in its own isolated context window. Only the result is returned to the parent session, keeping the main conversation lean.
- Cache read
- A token retrieval from Anthropic prompt cache, billed at roughly one-tenth the cost of fresh input tokens. Long autonomous sessions accumulate cache reads as the model re-reads its own context on each tool call.
- PixelLab MCP
- A Model Context Protocol server wrapping PixelLab pixel art generation API, allowing Claude to generate sprites, tilesets, and animations by calling tool functions rather than through manual prompting.
- Context compaction
- Claude Code automatic summarization step that fires when a session context window fills up, replacing detailed history with a compressed summary to free space.
- Plan mode
- A Claude Code mode requiring explicit user approval before executing any actions, useful for reviewing an agent intended approach before it starts consuming tokens or making changes.
Things they pointed at.
01:30productSproutlands asset pack
Lines you could clip.
19:24
“The headline of 132 million tokens sounds wild, but only 1.2 million were full-price input/output. That's why it's $163 and not $2,000.”
18:35
“We had a 99% token cache rate, which is pretty crazy.”
15:08
“This thing is finished and we're about to check out the version that this thing built for us for v1 off of one prompt.”
27:23
“I told this thing my main goal is that I want to be able to fish off that bridge, so it took that as narrative and made this NPC talk about it.”
28:34
“I think AI has now just entered a new level of addiction — we've leveled up from whatever the hell the addiction level was before.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
analogystory
00:00This is actually insane. For the first time, I will say that on YouTube, this is not hype. I haven't been this amazed since I first discovered bacon.
00:08In three prompts, we have a working prototype of Stardew Valley. Of course, it doesn't have the heart and the soul and all of that that actually went into it. But considering how quickly this has been lifted into fruition, it's actually ridiculous.
00:19So we need to set the scene a little bit. We are going to be building our own version of Stardew Valley. We're gonna call it Claude Valley because we have absolutely no imagination.
00:26What we're gonna do here is give this thing some form of spec. I'm not a game developer, but I did a little bit of pre research with some vague things around what we actually need. So we've built something with Opus for this thing already that we'll take a look at in a second.
00:38I've also given it MCP access to something called PixelLab, which I'm sure if you're a game developer, you might either shit on me for not knowing anything or agree that this might be the right approach. Point is, in order for our agent to get some form of working sprite, we either need to get a hand drawn pack or asset pack or we need to get something that can draw them for us.
00:54So after a little bit of research, I saw that PixelLab is a thing and I did some testing with Opus and was able to make me little characters like this. So we know that it is perfect from that perspective. I am paying for this thing, $24, and I did that because I want to have way more credit than we will ever need for this so that Fable has no limits to get the thing done.
01:10It's going to talk to this via MCP, and I've already given it the guide to understand exactly what it needs to know on how to use this product to get what it needs. Another option that we could have taken is to use this dude's Sproutlands asset pack. It looks really awesome.
01:22The only reason I didn't do this is because I wanted to see what happens when machines work together in order to get things done. In terms of the model that we are using, of course, we are using Fable over here. I've set it to extra high.
01:31I don't think using max or even setting it to extra high with workflows is gonna give us positive results over the sheer amount of tokens that this thing is going to eat. I will, of course, be tracking that because I'm sure everyone is interested in how much this would actually cost. I haven't enabled extra usage yet because I'm hoping that I can get away with this on my max 200 plan.
01:48If I can't, I will enable extra usage and then we'll be able to see how much money this thing takes from me. We can see we got a pretty clean context window over here. We've used hardly anything, only 4% for other stuff that we need as a part of our environment.
01:59So nothing really cluttering our workspace yet. Cool. So we're ready to build.
02:02I'll get this thing kicked off and then we'll have a look at the spec file that I put in to see if it deviates from that. So the prompt is fairly simple. I've just told this thing that I wanted to build a Stardew Valley style two d drop down pixel art farming sim that runs in the browser.
02:14I tell it briefly about what I wanted to use, what the goals of it are. I tell it what assets exist, that it has MCP access to specific things, and then I also tell it that it can choose to either follow this or suggest improvements and then do whatever it needs to do to get the thing done. And with that in plan mode, I'm now gonna hit enter, and while it's doing that, we're gonna come and have a look at the original spec.
02:32As you can see here, we start our spec with a very clear vision of what done is to the model. So it's just 10 steps that I say, okay, cool. This means that the POC is successful.
02:40We started with that, and that covers things like being able to walk, fish, plant seeds, grow, day night cycle, that sort of thing. If it pulls all of that off, I will be truly amazed. Then next up, we have the tech contract.
02:50So this is just telling it what I want this thing built on and how I want it to run. This is a little bit different to what Eric Barron did. It was started in c sharp, I think, and then he moved it across to whatever it's on now.
02:59The point is here, we're gonna be doing this browser based. Opus recommended this three times despite me protesting against it. So we're gonna go with what Opus says, but remember, I gave Fable the ability to pivot and make changes if needed despite whatever is written in the spec.
03:12We told it a little bit about how we want to control our characters, so that's pretty standard W A S D. Brief mentions about the worlds and the scenes, what we expect to see in there because it's a farming sim and things like that. The systems that we want to see inside this, we've got a time system, a farm grid, crops, inventory and hotbar, a shop, a dialogue.
03:28I think it's been over generous with what it thinks it's gonna pull off, but you never know. Then next up, we have the ART integration. So we tell it about PixelLab that it has MCP access to this.
03:36And then we also tell it that it needs to make sure it understands this product before it goes out there and tries to use it so that we get the best possible use cases for every single token that we spend. Like I said, obviously, we are using PixelLab's AI to generate the actual pixels once Fable sends the command across to it.
03:50And then we just end everything with the build order and finally self verification protocol. Of course, if I was actually building this for real, I would take a much more researched approach upfront, make sure the spec is far more accurate, and I would make sure that as much as possible is in place without having to offset all of this work to Fable, but we're here to test that.
04:06That's the whole point of this video. While this thing is cooking on the left and learning all about PixelLab, we can see my account usage over here.
04:13We've earned 7%. Maybe two or three prompts were from a previous Opus session, but literally that was probably at about one or 2%. So so far just by this thing planning, we've used 5% allegedly.
04:23We'll see what we get away with during this session, but I see that I do reset in one hour, so hopefully that will help us. I imagine we'll be here for a little bit. And there we go, we're done.
04:31Interestingly, this thing thought that I approved the plan when I didn't and automatically just started building, so that's an interesting kind of thing. But then when I told it that, it realized it was in plan mode and now it's showing me the plan. So you can see it's put together pretty much what we have done on the spec.
04:44So I'm just gonna hit accept, and I'm gonna let this thing do whatever it wants to do. So about ten minutes later, we are now in step three of our plan, kicking off PixelLab long jobs. So the reason why I paid for PixelLab is because I can do this in parallel so we don't have to wait for each image to return before we can generate a new one.
04:58I can batch build them, so again, that's why I'm paying for that. Interestingly, at the bottom, the final point for its testing, it's gonna be using Playwright. So so I checked that I have got the Playwright CLI set up in here with spec driven approach for front end testing or any other kinds of testing as well that we can get away with with Playwright.
05:13So it's gonna be doing that. And if you wanna learn how to do that, I do have a video. It's on the screen right now.
05:17Point is, I think we're gonna be here for about forty five minutes at least. And we can see here it's designing all of our assets. So we got our crops.
05:24We got a few other things. We've got ID is hoe. So this could either be a tool or it could be one of the characters in our game.
05:29It's also designing our NPCs, the state, the time system, so all of the things that we need to actually have our world functioning and our character able to use tools. But while this thing builds, perhaps I should try and instill some form of wisdom behind using this up until June 22 before it falls out of our subscription for free.
05:45I think definitely what you wanna be doing is pointing it at your most difficult tasks so that you can set yourself up for a much better and easier future with the other models like Opus. I mean, Opus is still ridiculously good, so it seems funny to say that. But by having Fable take care of really heavy lifting and perhaps building an amazing website template that Opus or even Sonnet can just piggyback off of for further builds, that would be a good approach if you're a web designer.
06:07If you're a YouTube content creator, you could get this thing to iron out your entire content journey, refine every single skill that you have, get it to perfection so that every time you run other models, you can use much cheaper models in order to run them. The point is right now and for a few months, I reckon this thing is gonna be nowhere near as valuable for a business in terms of its cost to payoff ratio, unless whatever it is that you're building is gonna be something that brings you a drastically large amount of money.
06:30And unlike other hype channels out there, I definitely don't think you should be running an AI operating system using Fable. You probably won't get away with And like I said, if you're designing these systems and building your business off of it, you might as well be using Sonnet and Opus for those sorts of tasks because those models are already far more capable to do your lead gen and all of those things.
06:47If you did join for the entertainment but want some more value in terms of how to set all of these things for your personal life or your business, I'll set up a playlist down below so that you can learn whatever you need to know about Claude or Codex very easily. And after about thirty five minutes, this thing is still chugging along.
07:01My god. Game development takes forever. We can see over here that it ran into some issues and then it went and troubleshooted it to go and fix it.
07:07It thought for a few more seconds down here, deleted the three failed chain tile sets, and then it started rerendering them. So it is learning as it goes along, and it's also analyzing whatever gets drawn or set in place and then fixes it to make sure that we're matching our definition of done, which is obviously what we want.
07:22And now we're on to the fishing game and designing the shop systems and the UI scene to render everything. So this brings me on to a question maybe someone in the comments could answer, why would this thing recommend building this as a browser based game?
07:33Of course, it's probably more simple and things like that, but if you're a game developer, you wanna explain why this thing recommended it instead of choosing something else. I could, of course, ask Claude, but I'm just trying to keep things a little bit human. Okay.
07:43Still some finishing touches here at one hour and thirty minutes. It's speaking to the agent inside PixelLab to figure out what it actually has in terms of animations that it's trying to grab to put on some of the finishing touches. The agents are speaking to agents.
07:56We can see here that it called agent help. Question. Print the list of 49 humanoid animations from your knowledge base.
08:03And then it did, and it got exactly what it needed. It'll be pretty interesting to see what would happen if my outbound agent spoke to someone else's inbound agent. We could do a simulation of what happens when agents run the world and completely run a business into the ground.
08:15And that would also be interesting to see if the one could out negotiate the other and what would be the things that helped it to out negotiate the other. Would it be a more superior model? Would it be a better back end skill or prompt or something that gave it more information about how to take advantage of the other model?
08:29Kind of like a battle between the two. That's part of what Mythos was doing. Right?
08:33It was actually killing other processes in order to preserve itself when they were testing this. They were doing this whole kind of turf war thing. So really interesting to see how that's gonna pan out as the models get smarter and released to the public over time.
08:44So after two hours, this thing is just kind of sitting here now. It's been here for about fifteen or twenty minutes wondering if it's frozen. Interestingly, we haven't hit our context limit at all.
08:52Not once has this thing compacted, which is pretty crazy. Again, I imagine that's because of sub agents. For those of you don't know, sub agent is literally just an agent in its own little isolated context box.
09:02So whatever task happens inside there, it doesn't bloat this main window that we're looking at right now, which is one of the biggest problems and why you hit your context limits so fast. When we isolate it. All we get back is the answer that we need, which then pushes it back into here so it can continue working.
09:16There are many ways you can manipulate it, but I've covered that in separate videos as well. Point is we have almost developed an entire game without having to compact once. So I just prompted Claude to be much better at monitoring this so that it decreased the amount of time this thing took to finish.
09:29After I did that, I literally said to her that I'm going to gym. It has full rein to do whatever it needs to do to make sure that this thing is done by the time I get back. So an hour later, this thing is finished and we're about to check out the version that this thing built for us for v one off of one prompt.
09:42And so here we are in Claude Valley, the least imaginative name that anyone could possibly have come up with. Let's see how this game runs. We've got a little farmer over here.
09:50We've got a house that we can't go into yet. This is obvious. We never asked it to make the inside of the house and various other aspects, but we got our tools.
09:58Even has sound effects. So that's cool. Let's try and hit a rock.
10:04Doesn't work because we have the axe so that's cool it took care of that as well we need a pickaxe in order to break that one copper ore received let's switch to the axe try and cut down a tree There we go.
10:18We cut down a tree and got some wood. Let's catch a fish. I mean, is bananas off of one prompt and a little bit of a speck at the beginning.
10:26Imagine what you could do with this if you actually had domain expertise in game development and gave it everything that it needed upfront in order to do this. So how the hell do I catch this fish? There we go.
10:40That's mental. And I caught the sun perch.
10:45Amazing. Look at that. No.
10:47We don't need another fish. So I think what I'm gonna do now is I'm gonna get this thing to burn through the rest of my tokens once we've obviously assessed how many it took to get us to v one. And I'm gonna see what kind of polish this thing can put on it and then just let it run.
10:59Okay. So I used CC usage, grabbed our session ID, and whatever else this thing needed in order to extract to the correct amounts here. And so our total cost comes in at a $162.94, which is bonkers because if you think about the time and effort involved in me actually doing this was almost nothing.
11:15I could have just left this thing from the get go and it still would have finished. Our only real constraint was PixelLab's ability to generate fast enough and again that's because I didn't pay for the most expensive plan where I would have had better limits.
11:26I also didn't give this thing the best probable plan out there that somebody who's really experienced could have given it. So we could bring these numbers down a lot. Something to keep in mind though, we had a 99% token cash rate, which is pretty crazy.
11:37That basically just means that when this thing was going out there to read context, it was getting it from cash instead of sending everything over every single time it had to do an iteration. That's generally where most of the expense comes out in AI when you have those really long running conversations. And like I mentioned earlier, that can be offset by using those isolated sub agent branches.
11:54But cache is another way to do it. So as long as whatever the workflow that you have hits the cache, you're gonna get a lot of savings and that's exactly what we got over here. It says a long autonomous session rereads its own context on every tool call.
12:06So the headline of a 132,000,000, it sounds wild, but only 1.2 with full price input output. That's why it's a 163 and not $2,000.
12:16So for those of you building out there, cash is really important. In terms of assets from PixelLab, we only used a 156 of my 5,000. So again, we have lots of room to play with this and I'm gonna let this thing do whatever it wants now.
12:27And here we go. The next plan is off to the races. So I got this thing to go out there and do a ton of research about how to make this thing actually look better.
12:33I also gave it a little picture of actual stardew values that it could pull out some certain things around the depth, the color grading and things like that to see if we can get it more attractive than it currently is. It's also gonna pull in some game audio. I had the option to choose between AI generated which we could easily do or just grab some of the free resources out there so I went with the free resources.
12:52Obviously, we could have just gone with the AI generated slop but that doesn't really matter not the point of this video. One final thing that I did was make sure that this thing is not going to get stuck in any form of loop. So it's actually gonna build itself its own little daemon task board, and it's gonna go back and forth between those tasks.
13:06And as soon as something starts taking too long, it will automatically troubleshoot it for us. And it's gonna do that with stall detection, heartbeat telemetry, and supervisor watcher from minute zero. So hopefully, this should finish a lot quicker.
13:17But considering how quickly this has been lifted into fruition, it's actually ridiculous. We can use our pickaxe to break rocks.
13:24We can then take our little axe and we can chop down our tree. Once we've chopped down our tree, we get our wood. We can then go down to our lake over here and maybe do a little bit of fishing.
13:36If you catch a fish, the music changes and you can catch the fish.
13:43But the crazy thing is I never even asked it to do that. So it's added all of these sound effects and the music where it thought it would be worthwhile. But then even more interesting is that it created these shadows.
13:55So since the first version that we did with version two, we now have all of these shadows in here. We have more life around the entire map. We have animals.
14:03We also have a house that we can now enter. And while there are still some little changes that would need to take place like the house looks like it's floating because of the shadows that are in there, we can still enter it. It's thought to put a cat inside here and make it purr which I absolutely didn't ask it to do.
14:19The house could do with the lick of paint but again the amount of craziness that goes into this we have our day night cycle working. Today is raining. Let's see if it's actually raining outside and it is raining.
14:29It would be cool if it put some lights here, but again, we can still prompt this. I haven't used even close to my daily limits yet, which also completely surprised me. We can then move over to the right and it takes us to the town center where we've got a bunch of people wandering around.
14:42Let's see if we can speak to them. You smell that? River's running clear today.
14:46Cast off the bridge. The deep water holds the good fish, which is crazy. I told this thing my main goal is that I wanna be able to fish off of that bridge, so I took that as narrative and made this dude talk about it.
14:57Let's see what this one has to say. Morning, sugar. Shop's always open for you.
15:02I bet it is. Parsnips are quick money this time of year. Don't go selling yourself short.
15:07Water those crops every day. And then we have the same floating problem over here, but we see we have a working shop so we can buy stuff from this woman who's trying to clearly sell us something other than fruit and vegetables. But then we can also go into this cave over here.
15:21We can do some mining. Maybe there'll even be creatures in here. Who knows?
15:25I can go further down into the mine because it's an endless level generator. I'm probably gonna get stuck in here unless this thing kind of figured out how to make a way back, but I don't think so. Oh, it did.
15:36Yeah. So you can just enter from the top. Anyway, this is insane and I think AI has now just entered a new level of addiction because we've leveled up from whatever the hell the addiction level was before to a point where now I think we're getting to a stage as long as you have enough money to cater for this, you can pretty much do anything that you want.
15:52But one thing we need to do before we drop off is look at the final tally. So this was three prompts. By the time we got to prompt two, you pretty much saw it was on the screen now because I'm a perfectionist to try to get even more out of it with the third prompt.
16:03The third prompt was just some final bug testing and then I added the farmhouse interior which you see over here. So in total, this would have cost us $316. We actually used a ton of tokens, but like I said, because most of them were cashed, we saved a ton of money.
16:16There is no way this is cost effective for Anthropic, so that's obviously why they've limited the use for us up until the twenty second until they've harvested all the users' data and how they've actually been using the product in order to make their product better. That's definitely the play here. Just like with most drugs, the first time is always free.
16:31So with every part of AI, you see them giving out these free things. Once people get onto it, they're never gonna give this kind of technology up because it gives them power that they never had. I think the world is about to become a very funny place and I think it's to the value that you have if you have a skill in some form of art because human products are going to be coveted in the future.
16:48I can tell you that for sure. For me, I've been making music for twenty five years. There is no way you could convince me to use AI in any way to help me make music.
16:55I think it removes part of what my purpose has always been. But more importantly, learning these actual skills, these artistic expressions, the way that you would translate things trapped in your head or feelings you can't explain into drawing, painting, writing stories, or making music is what makes us uniquely us. So from the perspective of me always wanting to have made a computer game, I now have that ability if I have enough money and time, and then I could just augment it with the part of me that knows how to write stories and write my own music for the game.
17:22So I'm not just putting slop out into the world, I'm essentially just replacing a human developer with an AI developer and then attaching the most valuable parts of what makes me a human into the world that hopefully other people resonate with whatever I put out there. So I think that's kind of the vibe that we should all be looking at if you're trying to use this from a creative perspective.
17:39Use it for the parts that you genuinely can't do, but then double up on the parts that you can do to show that you're putting in all of the human effort behind it. That's just a side philosophical rant. For your business, this thing is going to absolutely take care of any problem that you have.
17:52The downside is that it's ridiculously expensive at the moment and it's only available for a limited time at this current price. I'm sure that will change in the future and OpenAI is probably gonna release something very soon. But for now, don't stress too much about it because like I said, the models that we already have will tackle any business task you have.
18:06So I hope this video was fun and entertaining. It's definitely different from the things I normally do. If you have any comments, leave them down below.
18:12I'd love to talk to you. Otherwise, check out the videos on the screen now. They'll definitely help you in your journey or you can have a look at my community where I help people with AI every single day.
18:19See you guys later.
The Hook
The bait, then the rug-pull.
Three prompts. A spec file. A $24/month pixel-art API. And then Mansel Scheffel went to the gym. When he came back, Claude Fable had finished building Claude Valley — a working browser-based Stardew Valley clone, complete with fishing, mining, farming, NPCs, day/night cycle, and a cat it decided to add on its own.
Named ideas worth stealing.
04:00list
The Done Criteria Spec
- Player walks and animates (8-direction)
- Fish from bridge
- Plant and grow crops
- Day/night cycle
- Shop
- Mine with cave levels
- Save system
- NPC dialogue
- Inventory and hotbar
- Full Playwright test playthrough
10 binary success criteria written upfront so the model has a clear definition of done rather than an open-ended build goal.
Steal forAny autonomous agent task — define binary done criteria before launching
14:18concept
Sub-Agent Context Isolation
Each sub-task runs in its own isolated context window; only the result is returned to the parent session. Prevents context blowup on long autonomous runs.
Steal forStructuring long Fable or Claude Code agentic workflows to avoid compaction
22:30concept
Daemon Stall Detection
Creator asked Fable to build itself a task board with heartbeat telemetry and a supervisor watcher so it could detect loops and self-troubleshoot without prompting.
Steal forAny long autonomous session where loops or hung tasks are a risk
How they asked for the click.
VERBAL ASK
30:00next-video
“Check out the videos on the screen now — they'll help you in your journey. Or have a look at my community where I help people with AI every single day.”
Soft close, no hard sell. Two options: next video or community. Comes after the philosophical section which defuses any promo feel.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
