20:17David Ondrej · Tutorial
Hermes Agent is crazy… 180,000+ github stars
How MiniMax M3 sparse-attention architecture makes always-on autonomous agents 10–100x cheaper than running Opus or GPT-5.
June 8thThe argument in one line.
Fable 5 is the most capable Claude model yet, but the Claude app and raw API both suppress its output — the only way to unlock its full capability is to run it inside an agentic harness with built-in fallback logic like Cursor or Claude Code.
Read if. Skip if.
READ IF YOU ARE…
- You already use Claude, Cursor, or another LLM tool daily for coding or building and want to know whether Fable 5 changes your setup.
- You are choosing between API credits and subscription plans and want the real cost math before June 23 pricing changes.
- You want to see a live multi-agent build session with raw multi-threaded agentic work including errors and pivots.
- You build internal tools or solo side-projects and want a git auto-commit workflow that removes manual push friction.
SKIP IF…
- You want benchmark numbers or technical model comparisons — this is a workflow video, not a model evaluation.
- You do not write code or build software; almost all tactical advice assumes a coding or agentic-building context.
The full version, fast.
Fable 5 is Claude's first Mythos-class model, and it trips its own safety guardrails more aggressively than prior Claude versions. The Claude.ai app makes this worse by adding consumer-grade restrictions on top. The fix is running Fable inside Cursor's agent view or Claude Code, both of which fall back gracefully to Opus 4.8 when refusals fire rather than throwing hard errors. Subscription plans are 8-10x cheaper than raw API for heavy users, but Anthropic confirmed Fable 5 leaves subscription plans June 23. The video demonstrates these lessons live, building two parallel MVPs in a single session.
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 – 01:20
01 · Why the API is the wrong starting point
Refusals, no fallback logic, double cost of Opus — three reasons to avoid raw API for Fable 5.
01:20 – 03:30
02 · Best harnesses: Cursor vs Claude app vs API
Cursor agent view and Claude Code have built-in Opus fallback. The Claude app has extra consumer guardrails. API throws hard errors.
03:30 – 06:13
03 · Capability demo: one-shot 3D builds
Backrooms game, packet-as-cars visualization, Lovable app builder clone — all one or two prompts in Claude artifacts.
06:13 – 08:23
04 · Guardrail avoidance tactics
What triggers refusals (distillation language, security keywords), the XML-tag relay technique, rerouting through Opus first.
08:23 – 10:02
05 · Claude Code cost efficiency; June 23 deadline
Subscription gives 8-10x more usage than API for same spend. Fable leaves subscription plans June 23.
10:02 – 12:23
06 · Real AI spending and when to use Opus Fast
6k per month on OpenRouter before Fable. Opus 4.8 Fast same cost, dramatically faster — use it for speed-first agents.
12:23 – 14:14
07 · Model personality and startup idea analysis
First model that pushes back and calls out patterns. Used to evaluate 50-plus startup ideas simulating respected founder perspectives.
14:14 – 17:05
08 · Live build: Project Token Spend Tracker
Fable interviews the creator until it understands the idea 95%. AutoGit installed live. Two parallel agent threads running simultaneously.
17:05 – 21:00
09 · Prompting philosophy and agents.md pattern
Commission, don't micromanage. Create agents.md in every project. Mid-prompt sending while still typing demonstrated.
21:00 – 25:30
10 · Live build: YouTube-to-Twitter auto-reposter
Second MVP started simultaneously. GitHub CLI creates private repo without touching the dashboard. AutoGit self-pushes its own updates.
25:30 – 28:30
11 · Consolidated power-user tips from 3 deep research reports
Cross-report consensus: run in harness, commission don't micromanage, delete verbose scaffolding, add scope guards, handle HTTP 200 refusals explicitly.
28:30 – 32:29
12 · 3D simulations, traffic city, final CTA
City traffic simulator, fluid dynamics, 3D shooter — all one-shot. Closing: build something now, stop making excuses.
Lines worth screenshotting.
- The Claude.ai app is the worst place to use Fable 5 — it layers consumer guardrails on top of the model's already-aggressive safety filters.
- Cursor and Claude Code are the top two harnesses because both fall back silently to Opus 4.8 instead of throwing hard errors when Fable refusals fire.
- Asking Fable to 'tell me everything about Fable 5' is enough to trigger a refusal — the model reads it as a distillation attempt.
- XML tag workaround: run a sensitive query with Opus or GPT first, paste the output into XML tags, then ask Fable to evaluate it — rarely trips guardrails.
- Claude Code subscription gives roughly 8-10x more Fable usage than raw API for the same dollar — but that ends June 23.
- Running AI agents without Fable-class models is like competing against someone with no computer — the productivity gap is that wide.
- Verbose scaffolding prompts built for weaker models actively degrade Fable output — shorter, goal-oriented prompts perform better.
- Fable pushes back on bad ideas and calls out patterns you don't see — the first model where the creator felt it was close to his own intelligence.
- AutoGit: one Node CLI that auto-commits and pushes on every agent turn, removing all manual git friction for solo and internal projects.
- The agents.md pattern: a single file per project that tells all AI agents how to format responses — more durable than prompt headers.
- The 6-10k per month OpenRouter spend shown in the video is before Fable — a signal of where serious AI builders are already spending.
- For speed-critical agents, Opus 4.8 Fast is the same cost as Fable 5 and dramatically faster — use Fable only where intelligence is the bottleneck.
- Fable can clone any software in two prompts — demonstrated with a full Lovable-clone app builder.
- Giving Fable ambitious goals with few constraints outperforms detailed specifications — the model is a better architect than most humans.
- Anthropic added mandatory 30-day data retention for Fable — enterprise clients cannot opt out, unlike previous API-only access.
The harness decides how much Fable 5 you actually get.
WHAT TO LEARN
Fable 5 is powerful enough to trip its own guardrails in the wrong environment — the model you deploy matters less than the harness you run it in.
- The Claude.ai app applies consumer-grade guardrails on top of the model itself — Fable 5 underperforms there relative to what it can do in an agent harness.
- Run it in a harness with fallback logic (Cursor agent view or Claude Code) so refusals degrade silently to Opus 4.8 instead of breaking your workflow.
- The XML tag relay is the clearest tactical fix for refusals on sensitive prompts: let a less-restricted model run the query first, then ask Fable to evaluate the output.
- Subscription plans are 8-10x cheaper than API credits for heavy usage — but that window closes June 23 when Fable 5 moves to credits-only.
- Verbose scaffolding prompts built for weaker models actively hurt Fable output; shorter, goal-oriented prompts with clear success criteria outperform detailed specifications.
- AutoGit removes the last friction in agentic sessions — every agent turn auto-commits and pushes, so version history is continuous without a single manual git command.
- The agents.md pattern (a project-level file for AI agent behavior rules) is more durable than repeating style instructions in every prompt or system message.
- For agents where speed matters more than reasoning depth, Opus 4.8 Fast is the same cost as Fable 5 and significantly faster — save Fable for architecture-level decisions.
Terms worth knowing.
- Fable 5
- The commercial name for claude-fable-5, Anthropic's first Mythos-class model — priced at double the cost of Opus and subject to more aggressive safety filters.
- Mythos class
- Anthropic's model tier for frontier reasoning models with extended context windows and long-horizon agentic capability, distinct from the Opus/Sonnet/Haiku lines.
- Guardrail refusal
- When Fable 5 declines a prompt by downgrading itself to Opus 4.8 or returning HTTP 200 with stop_reason refusal — happens more frequently than prior Claude models.
- Cursor agent view
- A VS Code-based IDE interface that runs multiple simultaneous AI agents in parallel threads within a single project workspace, similar in design to the Codex app.
- AutoGit
- An open-source Node CLI tool that hooks into Cursor and Claude Code to auto-commit and push changes to Git on every agent turn — designed for solo or internal projects.
- agents.md
- A project-level markdown file that defines AI agent response style and behavior rules, kept separate from the README.
- Effort level
- A Cursor/Claude Code setting (Low / Medium / High / Extra High / Max) that controls compute per turn — Extra High is the practical sweet spot; Max shows diminishing returns.
- Permanent underclass
- A term for builders who cannot afford frontier model API costs once subsidized subscription pricing ends — the creator argues the productivity gap between model-haves and have-nots is growing irreversibly.
Things they pointed at.
Lines you could clip.
05:00
“There is gonna be a bigger and bigger gap between people on the cutting edge with the best models and people who simply cannot afford them.”
12:34
“This is probably the first model that I feel like is close to my intelligence. With Fable, you don't have that feeling — it will push back when you have a bad idea.”
18:05
“This is literally the future. You're running multiple agents, switching between projects, turning your thoughts into reality.”
24:12
“Fable legit doesn't understand how powerful it is.”
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.
metaphor
00:00So I'm running Fable five inside of Cmax, inside of Kerastro's agent window, as well as inside of Claude.
00:06And while most people will tell you, oh, the benchmarks are crazy. This is the new model, which you can just get by reading the official Anthropic article or by asking Claude to explain fable five. In this video, I'm gonna take a completely different approach, and I'm gonna focus on my workflow because I've been testing this model nonstop.
00:23I slept like a couple of hours because this is perhaps the biggest improvement, the biggest moral jump since the release of g p d four. It's incredible.
00:31But, again, all of that is useless if you don't know how to use it, and that's gonna be the central focus of this video. So first of all, you might be tempted to use it through OpenRouter, which is what I did.
00:41That was a mistake. There's a few reasons. Right?
00:43First of all, it's a very expensive model. It's one of the most expensive models in the world. It's double the cost of Opus.
00:48But the reason it's a mistake is because the biggest issue when using Fable is that it can trip the safeguards, and it can automatically reject a good amount of prompts. A lot of people are reporting it rejecting normal stuff, and it happened to me as well.
01:01You can literally ask something like, tell me everything there is to know about fable five. And just a prompt like this can trip the safeguards, and it will downgrade itself to opus 4.8, which obviously we don't want. We wanna be using fables.
01:13So there is a lot of tricks that I developed over the last twelve to fourteen hours. And the first thing is not using it through API because you need to have this built in mechanism of downgrading.
01:23Right? That's why I think currently, the two best places to use it is inside of Crisser and inside of Cloth Code or the Cloth app is because these have the built in mechanism for when the error there it is. You can see.
01:35Switch to Opus 4.8 for no reason. I didn't do any bioweapons. I didn't do any hacking.
01:41So there is specific way to prompt it. And after I cannot change my mindset about how to prompt this model versus previous models, this happens to me way, way less. So first of all, I would advise you to actually avoid the Claw app.
01:52This is where most people are gonna be using And, obviously, like, it's convenient. Right? You can use your Anthropic subscription.
01:58It's a simple user interface. You have the effort level. You can switch it easily.
02:01The issue is here, it's way more limited. It has a lot of prompting around it, a lot of guardrails, and that's because this app is gonna be used by hundreds of millions of people.
02:10I mean, tens of millions at the bare minimum, but probably hundreds of millions since Clor is quickly becoming more popular than ChaiGBD. So Anthropic made this very limited in the app, so I would avoid that. So then the second thing is like, okay.
02:20I'm gonna use it in the API. Right? So for example, here, I'm running py agent inside of c mux.
02:25It's powered by Fable. And this is still really good for a lot of different workflows, and I'm gonna explain how I'm using it, but it has some issues. And the main issue is that it doesn't have the fallback.
02:34So when it trips the safeguard, it just throws errors, and it doesn't try again. It doesn't gently downgrade to Opus, so I wouldn't use it in the API either.
02:43So if you're planning to use it inside of Hermes, inside of OpenClaw, inside of Py, it's not the best for most usage. There is still use cases for it, but it's not the best.
02:51Currently, I think the best is inside of cursor. This is where I'm using it and specifically in the cursor agent window. So this is what the default one looks like, you know, highly similar to Versus Code.
03:00After all, it's a fork of Versus Code, but this one is more similar to the Codex app. It's called agent window and or agent view, and this is how you can run many, many agents at once. So I've done a lot with this model over the past fourteen hours.
03:13I've built a new tool, I open source, which I'm gonna show you later on, called AutoGit, and I also released that to NPM. So I built it for me me and my team, and I also open sourced it. I also reviewed all areas of my life with Fable.
03:25I completely revamped some of our internal software to make it more stable, better designed, more reliable, faster. This is incredible model.
03:32Okay? Let me just say that. It's incredible model, and I'm gonna show you a lot of different things it can do in multiple different areas of life and business.
03:39Before I go into my workflows, let me show you a little taste of what Claude can cook up with Fable five. This is like a three d graphics game for, you know, backroom Slenderman. Look at the animation.
03:51Look at the feel. This is built all inside of cloth artifact. So this is not a full stack web app.
03:55This is not inside of an IDE. This is just a cloth artifact. Here's a second example of the type of games and assets Fable can cook up.
04:04This is a three d app where Fable five logs network packets and displays them as cars. So this is not just some pointless animation. These are the actual packets going to this computer rendered as three d cars in a nice cool animation.
04:17And stuff like this, Fable can one shot without any issues. Alright. So let me show you how I actually use Fable currently.
04:23And as I said, most of my usage is currently happening in cursor in the agent's view. So what I do is, you know, I work on a project. So for example, here is our internal software.
04:33I just create a new chat, a new thread. I select Fable. I have it default.
04:37And I either do extra high or I do max. Obviously, this is gonna take a lot of credits and cost cost a lot of money. But I'm gonna be honest with you guys.
04:46If you aren't prepared to spend multiple thousands of dollars per month for AI, you're not gonna have an advantage. Like, the models are becoming more expensive, and there is gonna be a bigger and bigger gap between people on the cutting edge with the best models who can run them for older agents and people who simply cannot afford them.
05:02So some people call this the permanent other class. You can call it whatever you want, but that is the harsh reality. Now for a new project, something like Autogit, which again, I just build that in a couple of hours with Fable, helped me get clarity about new project I wanna build, and then I basically described what I know about it and what I'm not sure about, and it started, like, interviewing me.
05:20Right? This model really puts in a lot more effort. It writes a lot more tests.
05:24It is competent in terms of when it makes a change, it doesn't just make it and push it. It tries to render the browser, tries to see what it looks like, tries to test it in multiple different ways to make sure it didn't break other things, that it works the way it intended, that it looks good in the front end. It really is like, when it was refactoring, for example, the canvas here, previous models were struggling with making this easy, especially when you have a lot of generations.
05:46Now we have only, like, eight images generating, but once you have 50 or 100 or diff on a single canvas, other models, including GBD 5.5 extra high, were struggling with this, and they were struggling to fix some of the bugs we had here.
06:01Fable just one shots them. It is expensive. You know, so be prepared to burn for your limits, but it is really, really good.
06:08Now let's talk about how to avoid the guardrails and not have fable rejecting unnecessarily. First of all, you don't wanna talk about any, like, cybersecurity or anything that looks like moral distillation.
06:19Distillation. Right? So even before, when I asked it, tell me everything there is to know about fable five, this is enough to make it think like I'm some Chinese lab trying to distill it.
06:28Right? So just be careful. Okay?
06:30Be careful by not making it sure that you wanna know everything about the model. Don't prompt it in a way that seems like you're building a dataset out of its responses for distillation.
06:39None of that. Second thing is cybersecurity. So if you ask for your app, like, check my app for security, sometimes it can reject that as well.
06:48What I did inside of a cursor here, when I was improving the security for YouTube alpha, what I did is I first told it to analyze the middleware authentication, all that to see if if, you know, all endpoints are properly secured, if someone who found an account could somehow get access, just basic stuff. It actually tripped the safeguards on that, and I think Opus ended up finishing it, the analysis, and then giving me a verdict.
07:10So what I did is I copied that verdict, and I gave it to Fable five, one million, context window max, put it into XML tags, take the output even if it's from Opus or GBD 5.5, and then resend that to Fable in a different window, tell it what do you think of this, you know, would you implement this in our code base, And, usually, that doesn't trigger the safeguards.
07:29The most cost efficient way to run it is obviously inside of cloth code because you're gonna get the subscription. So just make sure you do cloth update. Right?
07:38Cloth update for the latest version, and then launch cloth code, boom, and do slash model to switch to Fable five. There it is.
07:46You can do effort like usual. I recommend staying on high for most of you because otherwise, you're gonna run through limits.
07:54Extra high is, like, for most tasks. Max is max is really when you need the most.
08:00Right? Personally, I'm I'm running it on these two modes, but really the jump from x high to max is not that substantial. So if you want, you know, really, really good outputs, but you don't wanna be unnecessarily burning, stay on extra extra high.
08:14For quick chatting with it inside of, you know, Cloth app, stay on high. So Cloth code is gonna be the most cost efficient way to use it because the Anthropic subscription is subsidized. What that means is that you get way more usage than you would with API credits.
08:27So if you pay a $100 a month for the ClotheMax plan or $200 a month for the highest ClotheMax plan, you're probably getting, like, $2,000 worth of API credits. So if you go, for example, to OpenRouter and you wanna use it here inside of Py, like I do, you know, here, we can launch a second agent.
08:45Inside of ClothCode, you'll probably get roughly eight to 10 times more usage for that same $200 than you will get inside of Hermes agent, PyAgent, OpenClaw, anywhere where where you have to use the API pricing because it is not subsidized. Now on that context, I think Anthropic said that subscription they're gonna remove it from the subscriptions.
09:06Right? Here it is. On June 23, we'll remove Fable five from those plans.
09:11So it will only be available with credits with API credits. So that's bad news for anybody who is running multiple cloud subscriptions to take advantage of these subsidies.
09:22And, really, this is why people are saying now is the moment when the permanent underclass is starting to happen. Because now, sure, you maybe you're able to pay a $100 a month for a cloud subscription and you can, you know, chat with Fable here. No problems.
09:36Maybe you're even paying $200 a month. But the moment you have to switch to API credits and you have to pay the full price for every single prompt, your costs are gonna skyrocket. And I know that because I'm speaking from experience.
09:48I can show you my open router spendings here. So this is just the past one week. You can see 1 k spent.
09:54If we go just past one month, I'm all at 6 k spent. Right?
09:59And this is before Fable was even released. So imagine how this will change now when I'm switching most of my agents to Fable.
10:07I'm still using Opus 4.8 fast, by the way, because this model is the same cost as Fable and it's way, way, way faster. So for the agents where I need speed and I need a highly competent model, I'm still gonna be using this. For example, my calorie tracking agent in OpenCLOA, but for any agent where I need the best model, I'm gonna be switching to Fable.
10:26Right? So if I'm currently spending roughly 5 k, 6 k a month on OpenRouter alone, this is not even factoring my Clothe subscription.
10:34Right? Like, Clothe code subscription is not factored in here. Kerstor subscription is not factored in here.
10:39This is just OpenRouter. So my current AI spend is probably 8 to 10 k a month.
10:45Imagine how that's gonna change once I switch everything to Fable. Now another thing that Fable really excels at is three d models and creating three d graphics.
10:54So if you're working with Blender, Maya, or any other three d modeling software, this is by far the best model in the world at this. Yeah.
11:01This prompt, it was asked to create a three d robot, and it did that. One shot.
11:05No problem. That's gonna gonna be a theme. Right?
11:07Like, a lot of the apps you have, it gets just one shot without issues. In fact, Riley Brown had a good tweet on this where, basically, Fable can copy any software.
11:16Right? Any software you want, you give it, you describe it, it can just copy it.
11:21Now if you think about the implications of that but in this example, it just copied Lovable, which actually lets you build apps.
11:29Right? So it's not just a web app front end. This is a builder.
11:32It built a full app builder where it can create any app, and it did that in two prompts, not 20 prompts, not multiple weeks of development in two prompts. Okay? So the implications are any software can be cloned within couple of prompts, and that means it has never been more valuable to open source the software because a lot of people are not gonna be building with your software at all if it's not open source.
11:54And also the implications are you need to know how to use it, and you need to give it these ambitious prompts. Give it tasks that you previously wouldn't even consider giving to AI just because this model is completely different. You cannot treat it like the previous models.
12:09You need to give it more ambitious things, less constraints, less restrictions, and give it harder, bigger goals. And explain why you're building it and let it figure out the best solution for you instead of limiting it too much with too many details and specifications, where in fact, the model is probably smarter than you.
12:26I would say this is probably the first model that I feel like is close to my intelligence. Like, most of the models, when you talk to them, you can feel like, you know, they're a yes man. They're kinda agreeable.
12:34With Fable, you don't have that feeling. Like, it will push back when you have a bad idea. It will call out patterns in your messages that you don't see.
12:41Sometimes they're wrong. It's still not like super intelligence, definitely not, but there is a strong argument that this model is AGI.
12:49But also, I've used it for analyzing startup ideas. Right? So I'm getting the appetite to build something again, to launch another AI startup.
12:56If you have anything you want me to build, actually, feel free to comment below. Tell me what problems you have with the software you're using, you know, what type of apps you wish existed, what issues you're facing with the agents you're using, and maybe I'll pick your idea and build it. So if you have a clear start up idea you want me to build, make sure to comment below.
13:12But, anyways, I have more than enough of my ideas, and I used Fable to analyze them, pick the best ones, disqualify the worst ones, to analyze it through the optics of people that I respect. Right?
13:25Like people I had on my podcast, people on other podcasts, like Successful AI Founders, and review those ideas as if those people were looking at them directly. It's gonna be way better at that than Opus, Sonnet, GBT.
13:36It's a lot less agreeable, and it's a lot more objective. So here I asked it what to review my ideas. M d, which, you know, has a lot of things in there, maybe, like, 50 plus ideas.
13:46And I asked it which of these are small enough to build a quick MVP during a YouTube video and to give me top five candidates. So here they are. So let's see what we could actually build here.
13:55And, again, I'm inside of the cursor agents view in case you're not familiar with this app. It looks highly similar to the Codex app, so there's definitely some copying going on, but I think that's between everybody on the cutting edge of AI.
14:07So let's see. Auto repost to YouTube, poll channel download, post a caption, blue API grew demo able in one sitting. Okay.
14:14This is interesting. Put a project token spend tracker. Yeah.
14:17A lot of people are just vibe coding and they have no idea how much they're spending on that project. So let's go with this one. Let's go with this one.
14:25I'm gonna create a new folder for this. So we have this idea. Right?
14:28I'm gonna say idea. Boom. Idea.
14:33I want to build the following idea. So first of all, create a read me file and write a clear description of that. Keep it super concise.
14:40Do not add anything I didn't tell you. Boom.
14:44Let me start that off. Actually, this is a prompting tip. Do not write super long prompts because you can just send the prompt halfway through and already have the agent working on it while you're sending the second half.
14:54Right? Now I want you to interview me until you understand 95% about what this idea should look like.
15:03I can pre send this. There it is. And, actually, I can use my auto get in here.
15:08So here in the left sidebar inside of the cursor agents view, I can just go to a different thread here. And here's description of what AutoGit is.
15:17So I can just copy that. Boom. And I can launch a new thread in this new project.
15:22Say, set up AutoGit for this project. Boom. And I just paste in this.
15:29Honestly, this I I just overcomplicated that. The setup is way simpler. It's literally this.
15:34NPM install AutoGit, AutoGit setup, and then for the repo, we just do AutoGit on. So that's what I should have done. You can actually use the integrated terminal here.
15:45Auto git on. Okay. The it's there is no git repository, so I can actually turn it off here.
15:53Let me instead say that. First off, a git ignore and start a git repo in here.
16:00So we're gonna do that in this thread. In the second one, it's interviewing me. Whose YouTube channel are we polling?
16:06For the MVP, let's do my channel only. And I'm gonna give it the link to my YouTube channel here. Boom.
16:13But make it clear in the read me file that we wanna extend it for any YouTube channel in the future. Right? I answered that.
16:19Let me see if the Git repo is established. It is established. So now if we do auto git on, it should work.
16:27Auto git auto push on, every agent turn in this repo will push to git. Amazing. So now I never have to worry about git staging, git commits, or pushing ever again.
16:36And once let me make this clear. This is not ready for large apps with hundreds or thousands of paying users in production.
16:43This is for internal software. This is for, you know, hobbies, projects, your side projects that you want to build quickly, and you don't wanna be constantly pushing to GitHub or even telling the agents to push to GitHub because all of that slows you down. Right?
16:55So if you are an agentic engineer and you wanna move as fast as possible, definitely check out Autogit. It's fully open source. It's completely free, available on GitHub.
17:04I'm gonna link it below the video. Alright. So let's see.
17:07Here, is interviewing me with these questions. What gets reposted, Twitter caps, video length?
17:13That's not true at all. That is not true at all, actually. I don't know where you got that information.
17:17Boom. We can start that. Also, here, I'm gonna say create an agents dot m d file in this project, which is gonna explain how AI agents should behave in this project.
17:27The first sentence you should add is to make all your responses very concise and clear. There's a classic always create agents dot m d file in every single project.
17:36Now while we're going here, like, can literally switch to somewhere else, and we can, you know, do something. So I can say, what would be the next extensions for this?
17:44Analyze our code base, see what's currently implemented, and tell me what is the next sensible move. How could we make this even better, even more useful, even more intuitive? Be very concise.
17:55And this is literally the future. You're running multiple agents, switching between projects, speaking at the fastest speed you can, like, .
18:05This is a taste of what the future will look like. The time it takes for your idea to become reality has never been lower. It's all about a setup and you knowing how to use these AI models.
18:15Right? So if you can set this up, if you can set up cursor with, you know, multiple projects, with AutoGit, and you just have the setup where you can use Fable as efficiently as possible without running into token limits and without triggering the safeguards, and that's the setup side. And then if you can actually use it efficiently and clearly verbalize your thoughts and describe what you're trying to achieve, how you want it to look like, how it should be used, you can build anything.
18:38Like, the people with the best ideas and the best work ethic are gonna crush everybody else. I don't care if you have fifteen years of developer experience. You're gonna get destroyed by somebody who's an entrepreneur who can solve any problems because you really need that problem solving mindset.
18:52With Fable, there is no problem you cannot solve. Also, when it comes to simulations of stuff like three d fluid dynamics, Fable is significantly better than other ones.
19:02So these are other cutting edge models. Right? This is not smaller models.
19:05We have OPUS 4.8, Gemini 3.1, GBD 5.5. None of them is as realistic as Fable or detail oriented as Fable.
19:14They're not bad. These are still really good models, but if you can look at the wheels, right, the black hole, like, you can see this as qualitative improvement.
19:23This is not a small incremental tweak. This is significantly better at understanding the world.
19:28It has better understanding of the entire world. One thing it is limited on is AI or machine learning research on purpose, but other than that, it's basically better all across the board.
19:38Alright. Back to cursor. Let's see what's happening here.
19:40Understood. I have enough. What actually is okay.
19:44So I'm gonna say make your answer simpler and shorter. Let's see what's here. Alright.
19:48So, basically, I think I have a good clarity about this. Let's see. That's not true either or maybe the API.
19:56Okay. Maybe the API that's true. So long videos will be down at 07:20 p.
20:02Yes. Okay. We should implement a simple mechanic that if the video is over 512 megabytes, we should use FFmpeg to rerender it, to compress it to seven twenty p.
20:14And if that's still an issue, we should use a lower bit rate until it is below five twelve megabytes. So just make that clear in a very concise way so our program has these fallbacks.
20:27Okay. I think we should have enough clarity to start building this. So how do we make this happen?
20:32What do you need from me? Can we start running this locally? Can we test this locally before deploying this on a VPS, Or what are the next steps?
20:40Do you need me to set up the xAPI? Work work with me. You are my partner in this.
20:47Help me complete the MVP as quickly and efficiently as possible. And by the way, we should have the autogit here set up.
20:54So let me let me see if that's true. Okay. I think I activated autogit here.
20:58Let me know if it's working or not. Check the git commit history. Yes.
21:02It added this autogit undo, and it got automatically committed to the autogit repo. So that's kinda meta.
21:08You know, this there's a auto get running on the auto get repository, so it's self developing, I guess.
21:15I mean, not a self developing, but self pushing. You know? It's like a meta concept here.
21:18And then the second thing is readable comment messages. Now the comment message is the prompt you give to your agent. That's much better.
21:24Okay. So investigate how we installed it or how we set it up here on my MacBook for this project and then update it to the latest version.
21:33We're not pushing because the GitHub is not configured. Maybe we can configure that right now. Check if the GitHub CLI is configured on my MacBook.
21:43And if so, create a new GitHub repo with it and, yeah, push everything to GitHub.
21:51K? We can delegate this task to it. So let's do that.
21:58Open this up. Sign up for your x account, add $10 of credits, create project plus app. Alright.
22:05So it's gonna help me through the setup. Oh, by the way, I did more deep research on the best power user tips for Fable. So what I'm gonna do is I'm gonna do web or a text.
22:14Just paste all of this in here. Boom. I also have one from Perplexity Computer.
22:18Let me reload that. It's lagged. ChadGPT 5.5 pro extended.
22:23Let me paste that in. And we're gonna do this one as well. And, basically, I'm gonna have Fable summarize the results of these three deep researches to see what are the shared tips that power users have.
22:37Let's see what's happening here with the Twitter repost through one. Scroll down. Alright.
22:41So the GitHub CLI was configured, and it did push it. Let me click on that. Hopefully, it made it private.
22:47It did make it private. Okay. See, like, these agents, they can do so much.
22:50Like, I don't even have to go to GitHub, create a repo, just use the GitHub CLI to do that. I have AutoGit, which automatically pushes. AutoGit has now has a remote to push, so all that is good.
23:00Boom. Paste that screenshot. Here's what I see now.
23:03Tell me where to go next. Be very concise. K.
23:07What's let's look at the top power user tips. Okay? Consensus across all three reports.
23:13When to use it? Only for hard, long horizon work. I completely disagree.
23:17I completely disagree. You should use it always if you can afford to. If you can't afford to, then, you know, that makes sense, but you should use it as much as you can afford.
23:26Like, don't use it only for big tasks. You know?
23:30Use it anywhere where that intelligence could be squeezed, which I think is most tasks. Anyways, how to prompt it? Commissions don't micromanage.
23:38Okay. That's what I said. Give it a goal, some constraints, success criteria, and let it run.
23:43Stop trying to think you know the implementation better. These models are way too competent.
23:49They are very, very competent. They probably know better than you. They're probably better code based architect than you.
23:55They're probably better software engineer than you. Just describe what type of app or program you want, how the software should work, how it's gonna be used, and let it figure out the details.
24:06Delete old verbose scaffolding prompts. They now degrade output. Yeah.
24:09So all of your prompts that were, like, for less powerful models, they're just wasting tokens for Fable. You should be more vague while being very clear about what it should look like.
24:19Be more vague about the implementation. Add a scope guard. Simplest solution that works.
24:23No extra refactoring. It's very important. Agent loop.
24:26Run it in a harness. Yes. Completely agree.
24:29Cloth code or cursor, as I mentioned in this video, are the two best harnesses right now to run it. Not raw chat. Don't use it in Cloud dot AI.
24:37Guardrails and API plumbing. Refusals, HTTP 200, stop reason refusal, handle explicitly.
24:42So when you're building apps, like if you have an AI agent inside of your software, you need to handle this properly. Because of you know, for example, right now, using it inside of pie, it doesn't really handle it properly.
24:54That's why it's better to use it in cloth code or cursor. Oh, yeah. And this is important.
24:58Anthropic added mandatory 30 data retention. So and a lot of enterprise clients are not gonna use Fable because of this.
25:06I think a lot of their employees are gonna use it secretly. But what that means is basically Anthropic has to or has the option to look at their data for thirty days no matter if you use it in the API, no matter if you use it inside of cloud.ai.
25:21In the past, if you used it through the API, they didn't retain any data. But now with Paypal aka Mythos, they do look at everyone's data, and a lot of companies will not use it because of that.
25:33So we'll see if they change that or not. Alright. Let's look at what's happening here.
25:38It's guiding me more how to set it up. So let's do that. Gear icon, user authentication, read and write.
25:43Okay. So I'm getting this error from the developer API. I don't know if it's down or not.
25:48I'm getting this error when trying to access anything in the X developer dashboard. Let me check if we have agents.
25:56M d. We do. And that's not how it should be rewritten.
26:05Update agents. M d by removing any and all information about this project architecture that should be inside of read me .md.
26:15Instead, agents.md should be very focused on the response style of the agents, such as make your answers clear and very concise.
26:24It's I'm getting this here as well. That's crazy. Check the built in browser here inside of cursor.
26:30I'm still getting the same error. Alright. While that's running, let me show you a bit more insane stuff that people have built with Fable.
26:36It is a mechanical engineer task. It can generate intricate working assemblies and mechanism in a single prompt. Let's look at it.
26:45Pick up the cube. Damn.
26:50Guys, I I don't even know. Like, these models, they they can basically do anything. Like, really giving it the right harness, the right environment, and an ambitious enough of a task that you can just do it.
27:01Another thing that Fable just one shorted in a single prompt is this city's traffic simulator with all these boxes, all these three d properties, physics.
27:12One shorted, day night cycle. If you're not building stuff with AI, I don't know what to tell you. Like, I don't really care what you're building.
27:18You need to be building something. Internal software, an AI startup, some open source repositories.
27:24Just build something. Improve your skills. Get familiar with these tools, get your hands dirty with these tools.
27:31This is the future. The people who know how to control them and how to get the most out of these tools, they're gonna have such immense advantage.
27:39There's no even there's no way for me to put it into words even how big of advantage we have. Like, imagine you competing against somebody who cannot use electricity, cannot use Internet, cannot use computers.
27:51Like, you would destroy that person in business, in communications, in life, in fitness, in your health. You just completely destroy that person in anything, especially in business.
28:01Okay? Like, imagine competing to somebody who doesn't have a computer, doesn't have a smartphone, doesn't have the Internet, doesn't have spreadsheets, calculators, or electricity.
28:11You could crush that person so easily. That's gonna be 100 times less of a difference than somebody with access to the latest, greatest models who actually have the has the budget to use them and who has the competence and skill to get the most out of them.
28:25Okay. So this is kinda insane. While I was demoing the other things, Fable managed to fix the bug where this was not loading.
28:34You know, I was getting this error as I showed you. It managed to figure out if the issue is only with the deep links. And that basically, works.
28:41You know, some of these link redirects are broken, and it started setting this up by itself. I had to stop it because I got scared a bit because it was making too much progress, and this is my main Twitter account, I don't want it to do anything anything crazy.
28:56But it managed to figure out the bug that I couldn't do myself in two different browsers because the deep links were broken. It figured that out, and now it started setting this up. Man, this is this is a bit scary in a good way, though.
29:11Like, it it's a bit too good. Alright. So at the same time, I asked it to do a deep research on idea number four here and to give us a step by step plan, you know, basically analyzing comments and training a pain map, all of that.
29:26So, again, I'm gonna go in my channel and say, let's start just for my channel, and I need you to go into slash documents slash code on my MacBook, create a new folder there, and just fully build the MVP end to end completely and fully.
29:44All four phases. Yo. What is this?
29:47One to two weeks. No one to two weeks. We're doing it right now.
29:51These models, they never learn. You know? They always think this is, weeks or months.
29:55Week three to six. Now we're building this right now, Fable. I don't think Fable it's kinda crazy how like, Fable legit doesn't understand how powerful it is.
30:06It's kinda funny. Okay. Nothing saved yet.
30:09Both steps. Knees are okay. Fill the form permissions.
30:14Yes. Okay. This is another good example of what Fable can do.
30:18Here we have a three d shooter game. You know, if you played oh, wait.
30:23The bunny hop works as well.
30:31Oh, need to reload. These mechanics okay. The bunny hop is not as good.
30:36You know, I'm gonna try harding now. But this is built, I think, one shot. I would need a green crosshair.
30:40Anyways, this game is pretty cool. And, yeah, this is just a small taste of like, imagine everything that Fable can one shot.
30:49It can what can it two shot? You know?
30:51What can it three shot? Imagine doing 10 proms. Imagine actually putting some thought and effort and energy into it.
30:58Like, you you can build a full business faster than ever before. Like, really, if you if you don't think so, you're just full of limiting beliefs.
31:05So whether it's building mobile apps, video games, b to b SaaS, internal software, open source packages, I mean, just do do something, guys.
31:16Like, that's the main message. Do something. Stop making excuses.
31:19Fable is insane. I'm literally building two different MVPs right now. And, yeah, like, of the ideas I had on the list are just, like, one prompt away.
31:28So now whether it's, you know okay. You shouldn't have multiple main businesses.
31:33Right? So just pick one bet and put 80% of your time towards But next 20%, you should build something like tools for yourself, automations, things to speed up your workflows.
31:43Whatever it is whatever it is, just start building it. It's literally one prompt away. When you have a setup like this inside of the cursor agent's view, you could just start a new threat and launch it.
31:55So, like, build this end to end or do extensive research how we could build this and ask me the five most important questions about the biggest technical and software desired decisions. Something like that. Boom.
32:05It does the research, comes back fifteen minutes later, asks you a question. You answer them how it should be used, you know, how you imagine it, stuff like that, and then it just builds it. One or two prompts, and it's done.
32:15And, like, internal software has never been easier. External projects and business has never been easier. Man, it's gonna be crazy.
32:20The next one or two months, we're gonna see incredible renaissance of business, of products, of software. Things that were never possible are now possible.
The Hook
The bait, then the rug-pull.
Fourteen hours into Claude Fable 5's release, David Ondrej had already rebuilt internal software, open-sourced a new CLI tool, and spent a sleepless night discovering that the most powerful Claude model yet is also the easiest to accidentally break — and the title tells you where most people will go wrong.
Named ideas worth stealing.
15:30concept
AutoGit
Open-source Node CLI that hooks every Cursor/Claude Code agent turn and auto-commits and auto-pushes to Git. Install: npm install autogit, autogit setup, autogit on.
Steal forAny solo project or internal tool where you want zero-friction version history during agentic sessions
07:14model
XML Tag Relay
When Fable refuses a sensitive prompt, run it with Opus or GPT first, paste the full output into XML tags, and send that to Fable asking for its opinion. Rarely triggers guardrails.
Steal forSecurity analysis, competitive intelligence, any prompt category Fable refuses
17:30concept
agents.md pattern
A project-level markdown file defining agent response style and ground rules — separate from README. More durable than prepending style instructions to every prompt.
Steal forAny multi-agent project where different threads need consistent behavior without repeating system prompts
08:25model
Effort Level Selection
- Low
- Medium
- High
- Extra High
- Max
Extra High is the practical default. Max has diminishing returns vs xHigh. High for quick chatting.
Steal forCost management for heavy Cursor/Claude Code users
How they asked for the click.
VERBAL ASK
31:01subscribe
“Build something. Stop making excuses. Whether it is mobile apps, video games, B2B SaaS, internal software, open source packages — just do something.”
Passionate monologue close with no explicit subscribe ask — relies on energy and credibility established through the live session
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
