The argument in one line.
Fable 5 made step-by-step instructions obsolete — the only prompt structure that matters now is the GOAL framework: ground it in truth, state the outcome, grant autonomy over the path, and require proof it worked.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code and have noticed Fable 5 behaving differently from Opus or Sonnet — more autonomous, harder to predict.
- You have an existing project, codebase, or system you want to upgrade and want a repeatable prompt structure for handing it off.
- You want to use Claude Code dynamic workflows (UltraCode) but are not sure how to structure the initial prompt to get good results.
- You have been writing long step-by-step instruction prompts and finding the model sticks rigidly to them even when a better path exists.
SKIP IF…
- You are looking for a deep technical breakdown of Fable 5 pricing or context window specs — this is strategy, not spec.
- You have not tried Claude Code at all yet — the framework assumes you already have something to upgrade.
The full version, fast.
Fable 5 operates more like an autonomous agent than a chat model, which means the old habit of writing detailed step-by-step instructions actually hobbles it. The GOAL framework fixes this in four moves: point it at real source material before it reasons (Ground), define success in testable terms not vague directions (Outcome), let it choose its own implementation path (Autonomy), and build in checkpoints where it must prove the work rather than just declare it done (Loop in proof). The video demonstrates the full cycle live, building a content-card feature into an existing dashboard from a single prompt.
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 – 00:36
01 · Cold open + promise
Hook claim (most people using it wrong) + four-step GOAL framework promise + host intro
00:36 – 01:01
02 · AIOS dashboard demo
Tours existing Agentic OS — trend tracking, competitor monitoring, agent execution panel — establishes what is being upgraded
01:01 – 02:21
03 · How Fable 5 changed everything
Fable is a mythos-class model that fundamentally operates differently; best use case is upgrading existing systems, not starting from scratch
02:21 – 02:39
04 · Back to prompt engineering
The realization: we are back to prompt engineering but at the goal level, not the instruction level
02:39 – 04:27
05 · Paradigm shift: task vs. goal
Evolution from chat-era prompting to step-by-step instructions to single-goal prompts; the new question is whether Claude is doing the RIGHT work
04:27 – 04:27
06 · The live upgrade prompt
Reads the full live prompt aloud: AIOS folder + Obsidian vault + goal (content cards) + interview-me-if-fuzzy + autonomy grants
04:10 – 04:27
07 · Accessing Fable 5 + UltraCode
/model command to select Fable, /effort command to unlock UltraCode dynamic workflows
04:27 – 05:20
08 · How dynamic workflows work
Orchestrator spawns sub-agents (potentially hundreds), checker sub-agents verify the work — self-verification is the structural advantage
05:20 – 06:16
09 · G — Ground everything in truth
Make Claude read existing source material before reasoning; point it to real files, code, spreadsheets, Notion docs
06:16 – 07:27
10 · O — Outcome, not orders
State what done looks like in specific, testable terms; treat Claude like an employee and define success; be more ambitious
07:27 – 08:08
11 · A — Autonomy over the path
Give the destination, not the route; once you prescribe the path, Claude commits even if a better one exists
08:08 – 09:04
12 · L — Loop in proof
Require Claude to stop at big decisions, test with real data, and show before/after — freedom with verification checkpoints
09:04 – 11:57
13 · Free template + meta-prompt
Fill-in-the-blank GOAL template; meta-prompt strategy where Claude interviews you to extract your goal when you cannot articulate it yourself
11:57 – 12:33
14 · Live content cards built
Result: five real content cards from vault entries, each with hook, coaching lesson, format label, and three talking points; nightly regeneration via schedule
12:33 – 12:53
15 · CTA + outro
Community link (Claude Code Club, ), next-video suggestion
Lines worth screenshotting.
- The question for Fable 5 is no longer 'is Claude doing the work right?' but 'is Claude doing the right work?' — and only a goal-shaped prompt answers it.
- Step-by-step instructions backfire with Fable 5: the model commits to your plan even when a better path exists, so withholding the how unlocks better results.
- Make Claude read before it reasons — pointing it at existing files, code, or a knowledge base before you state the goal eliminates entire categories of clarifying questions.
- Vague goals are a tax: 'make it better' forces Claude to guess what you value, while specific criteria let the checker sub-agents actually verify the work.
- Dynamic workflows spawn hundreds of sub-agents that then spawn checker sub-agents — the self-verification layer is what makes a single-prompt upgrade reliable.
- Fable 5 costs roughly twice what Opus costs, so the economic case for getting the goal right the first time is stronger than with earlier models.
- If you cannot articulate your goal clearly, give Claude a meta-prompt that interviews you — it will ask one question at a time until the goal is specific enough to act on.
- Granting autonomy over layout and implementation is not laziness — it is a forcing function that pushes Claude toward the optimal path rather than your assumed one.
- Requiring a before/after comparison is a zero-cost quality gate: it surfaces regressions and gives you a clear accept/reject moment without micromanaging the build.
- A goal anchored to real source material (files, vault, spreadsheet) produces outputs grounded in actual context rather than hallucinated assumptions.
Four moves that let Claude work without a leash.
WHAT TO LEARN
Fable 5 does not need a recipe — it needs a destination, real source material to reason from, room to choose its own path, and a checkpoint that forces it to prove the work before calling it done.
- Before writing any goal prompt, point Claude at the existing thing you want changed — a folder, a codebase, a doc — and tell it to read the whole system first. This single habit eliminates most clarifying questions and anchors outputs to reality.
- Specific success criteria unlock self-verification: vague goals like 'make it better' have no pass/fail test, so Claude cannot check its own work; criteria like 'a visitor knows what I do in 5 seconds' can be tested against a browser.
- Granting explicit autonomy over layout and implementation decisions does not lower quality — it prevents Claude from locking onto your assumed approach and missing a better one.
- Build proof requirements into the prompt itself: ask for a before/after comparison and a browser test with real data, not a self-report that the task is complete.
- If you cannot write a specific goal, use a meta-prompt that instructs Claude to interview you one question at a time until the goal is concrete enough to produce testable criteria.
- Dynamic workflows (UltraCode in Claude Code) are most valuable when the task involves reading a large existing codebase and making coordinated changes across many files — the checker sub-agents catch regressions the builder agents introduce.
Terms worth knowing.
- Agentic OS (AIOS)
- A personal dashboard built in Claude Code that centralizes trend tracking, competitor monitoring, social channels, and agent execution — the system being upgraded in the live demo.
- Dynamic workflows
- A Claude Code feature (accessed via /effort, branded UltraCode) where an orchestrator agent spawns hundreds of sub-agents to parallelize work, with a separate checker layer to verify results.
- GOAL framework
- Four-part prompt structure: Ground everything in truth, Outcome not orders, Autonomy over the path, Loop in proof. Designed for goal-level prompting with Fable 5.
- Obsidian vault
- A local folder of interconnected Markdown notes used as a personal knowledge base — holds coaching call notes, audience language, and offer details that Claude reads before generating content ideas.
- Meta-prompt
- A prompt that instructs Claude to interview you rather than act immediately — used when you cannot yet articulate your goal clearly enough for Claude to execute against it.
- Testable criteria
- Specific, verifiable success conditions written into a prompt so checker sub-agents can confirm the work meets the goal rather than relying on self-reported completion.
Things they pointed at.
Lines you could clip.
03:15
“We've been asking, is Claude doing the work right? But now we have to ask ourselves, is Claude doing the right work?”
01:31
“One of the most overlooked use cases for Fable is upgrading things that are already working and making them 10 times better.”
07:37
“Just tell it where you want to get to, you need to resist telling it how to get there.”
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.
analogy
00:00Claude Fable five just dropped, but most people are using it wrong. In this video, I'm going to use Anthropic's most powerful model to upgrade my agentic OS with a single prompt and teach you the four step system to using Claude Fable better than 99% of people. And if we haven't met yet, my name is Duncan Rogoff.
00:17I'm a former art director for brands like Apple, PlayStation, and Nissan, and I now run one of the top communities for learning Quad Code and building income. By the end of this video, you'll know how to use Quad Fable to build exactly what you want every single time with out micromanaging it and without burning tokens.
00:33So focus in, close all your open tabs, and let's build. So this is my agentic OS or AIOS, and this is how I like to operate. Now it doesn't matter if you don't have one of these because most people have something they've already built that could use an upgrade.
00:46And so this is already pretty good. It allows me to track any of the trends in my niche, all the content my competitors are putting out, my own social channels, any of the projects that I'm working on, and I can even run any of my agents from right here inside of my dashboard. But I think it could be better.
01:01If you haven't heard, Quad Fable five just came out and it is a mythos class model and its capabilities exceed those of any model Anthropic has ever made. What I have found is that people are using Fable five to build some incredible things, but they're starting from scratch. And I think one of the most overlooked use cases for Fable is upgrading things that are already working and making them 10 times better.
01:22But Fable fundamentally operates different than any of the models that have come before it. Now our whole job is to simply give Fable a goal and let it execute. You can now run an entire project from a single prompt.
01:35And at first I couldn't figure out why this felt so new and wonderful, but also so familiar, and then it hit me. We are just back to prompt engineering. So two years ago, like when everybody was just in chat, prompt engineering was all the rage.
01:47How can you construct the most perfect prompt to get exactly what you want out of these chatbots? A lot of people were doing this in ChatGPT, but then the models got smart and the game completely changed.
01:58Instead of doing all this prompt engineering, it really became about writing these step by step instructions for the tasks you wanted these models to perform. For instance, go research some information about my audience, create a lead magnet for me, draft a LinkedIn post, publish it. But in order to get good results, you have to be pretty specific.
02:15But now Fable does the entire job from a single prompt, and so the question we are asking has completely changed. We've been asking, is Claude doing the work right? But now we have to ask ourselves, is Claude doing the right work?
02:28I'm gonna give Fable pretty much free range to upgrade my AIOS, which is the dashboard that runs my entire content business. And instead of giving specific instructions, I'm going to give it a goal, and this is the prompt I'm going to use.
02:39And in a little bit, I will give you a template that you can use if you don't have your own AIOS and even a meta prompt that instructs Claude to interview you to figure out what your goals really are. So you can literally be starting with nothing and this thing is gonna cook. So here's the prompt we're using today.
02:54Here's my AIOS in this folder and my Obsidian vault. Read the whole system first and map what it does today. So if you don't know what an Obsidian vault is, it's basically my second brain or my digital brain.
03:04I'll leave a video up here that shows you how to build one for yourself and this basically has all the knowledge about me, any of the coaching calls I've been on, any of my offers, my audience, things like that. This knows everything. And again, it's okay if you don't have this.
03:16The whole idea is that you are giving Fable the things you're hoping to upgrade. Here's my new goal. When I open my dashboard, it should deal me content cards I can shoot the same day.
03:25Each card is a topic idea with a written hook anchored to something real, a member win, a coaching lesson, or the exact language my audience uses. Turn this into testable criteria, interviewing me if anything is fuzzy, and check them with me before you build. All of that intelligence will have somewhere in my vault, explore it yourself and figure out where.
03:43And where the cards live in the UI, a new panel, a page, a slide over, it's your call. Basically how the things look. Make it fit the design system that's already there, then build it.
03:53Verify in the browser with real vault data and check-in with me before any big decisions. Finish by dealing me five real cards and showing me where each one came from. So this is the structure of a high quality Fable five prompt and I'm going to break down each one of these pieces for you, but let's kick Fable off and let's start our upgrade.
04:10And I just opened up a terminal window. I'm going to fire up Claude. And then in order to access Fable, it's super easy.
04:16All you have to do is type in slash model and you can come in here and now you can see you can select Fable. There's one other piece that makes us 10 times more powerful and it's another new feature from Anthropic called dynamic workflows. In order to access them, all you have to do is type in effort and you can see we have access to something called UltraCode which unlocks these insane workflows.
04:35So the first thing I'm going to do is I'm just going to give it my AIOS folder which basically has all the code for my operating system, and then I'm just going to paste in the prompt. So this is gonna start to get to work. And just so you understand what dynamic workflows are, basically you give a command to the orchestrator agent which sits at the top.
04:52It spawns up a series of sub agents. This could literally be hundreds of sub agents to complete the task. And then from there, it spawns up another group of like of sub sub agents to check the work of the agents that are actually doing things for you.
05:06We are now rapidly reaching this crazy level of AI where Quad is executing on these tasks and then determining itself whether or not it's doing a good job. So let me just break down the four step system for you so you understand how to craft a high quality prompt every time. And it comes down to the four letters, g o a l, goal.
05:24Our whole idea now is we are giving Claude a goal rather than a task. What is our ideal end state? Where are we trying to go?
05:32We can see here that Claude is coming back asking me if I want to run this dynamic workflow. It's going to spin up multiple sub agents to map all of the AIS CodeVault and my Obsidian Vault intelligence, and then it's going to fill in any of the gaps. Just so you do know that dynamic workflows are really expensive, the Fable model is about two times as expensive as Opus, so you do want to be careful about your limits when you are using this model.
05:55Just gonna say yes, run it, and we can see that this has spawned five different sub agents for us to do the work. Anyway, g stands for ground everything in truth, o is outcome not orders, a is autonomy over the path, and l is loop in proof. So g is we want to ground everything in truth, especially if you are looking for an upgrade.
06:14So make it read something before it reasons. If you have something that already exists, point Claude to it. Let it read the whole thing first.
06:21So like for example, like here is my personal website in this folder, read all of it first. You can give it an app that you built, a spreadsheet, your Notion system, anything that you So again, in my example, I said here's my AIOS in this folder and my Obsidian vault lives here. Next is o, outcome not orders.
06:38Tell it what does done look like. Treat Claude like an employee. What does success look like to you?
06:44Be specific, don't be vague. Don't just say, make it better. Say, I wanna go from idea to a finished post in ten minutes.
06:51If you are unable to say clearly what your goal is, tell Claude to interview you to figure it out. Now you wanna be treating Claude like a thought partner, give it goals, and because of these new models, you can be way more ambitious. A goal might be something like when somebody comes to my site, a visitor knows what I do in five seconds and books a call.
07:10Or I wanna make sure that everything I post has a strong hook or that drafts sound like me or that something lands every single morning. For me, I said deal me content cards that I can shoot the same day, each anchor to something real. So my whole thinking is that can I get this system to surface brand new ideas for me to create videos about?
07:27Next is a, it's autonomy over the path. Again, give the goal not the steps. These models are so smart, they're going to take a path that you might not even understand.
07:37Just tell it where you want to get to, you need to resist telling it how to get there. Because once you tell it how to get there, it's going to stick to the plan, but it may not be the best path. So let Claude pick the route and it will often find a better way than you would.
07:50For example, figure out the layout and the words yourself. Don't ask me. It can decide which files to change, which tools to use, how to build it, which order to go in, etcetera.
07:58I said, explore everything yourself, figure out where to put stuff, and where the cards live, you decide. And the last piece is l, make it prove that it worked. Make sure that it is checking itself.
08:09So it's great that it has freedom, but you don't wanna give it too much freedom. If there are any important moments that you want to have involvement in, like tell it to stop at the big decisions and ask you a question first. Tell it to test the work, not just say it's done, and then have it show you the before and after so you can compare.
08:24So for example, if you're building a website, say build it, open it in the browser, show me before and after, and check with me before any big change. So for the website example, here's how you would use it. Instead of just saying like, hey, make my website better and make it look more professional, you would say something like, here's my personal website, read all of it first.
08:41Your new goal is that a visitor knows what I do in five seconds and they book a call. Figure out the layout and the words yourself, don't ask me how, build it, open it in the browser, and show me before and after, check with me before any big change. So this is how to use the goal framework to craft a high quality prompt so Fable will execute exactly as you want it to without you having to micromanage it and without you burning a lot of extra tokens.
09:04So here's a template that you can use. You just have to fill in the blanks like the thing that you already have and what your goal might be. You wanna get access to this prompt and the meta prompt where Claude interviews you, I'll leave a link in the description.
09:15You can get this totally for free. But the meta prompt is super cool and I recommend everybody use this strategy of just having Claude interview you because you can, like, ramble and brain dump anything that you want, but you still might miss stuff, and Claude may still have questions. So you can actually go one level up.
09:29If if you're not sure what your goal is or what you're starting with, just have Claude use this prompt. Just say, I have a project for you, but I'm not ready to brief you, so interview me first. Ask me one question at a time in plain language.
09:41You're trying to learn four things. G, what I already have, files, a folder, a site, a system, a doc. Get the location then read all of it yourself before asking anything it could answer for you.
09:53So even in this step alone, you're giving Cloud access to a whole bunch of information and it may save it from having to ask you any more questions. The next question is ask me what do I want it to become? Push past vague answers.
10:04If I say make it better, ask me what better looks like and how we both know it happened. Turn my answers into specific testable criteria.
10:12This is how those checker agents can actually verify the work of the sub agents. Figure out what's mine versus yours, find out which decisions I actually care about. Anything I don't claim is your call.
10:23And last, what proof do I need as the human to know that something is good and right? Ask how I want it verified when you're done. Do I want it open?
10:30Do I want it tested with real data? Like, at what point am I happy with this? Keep the interview short, five to seven questions, then stop.
10:36Make reasonable calls on the small stuff yourself, then write everything back to me as one master prompt. When I approve it, execute it and check-in with me before any big or hard to undo decisions. So we are now at the interview stage, and Claude is asking me a couple of questions.
10:51So what format should the cards target? Shoot the same day reads short form to me, but coaching lessons also make strong long form topics. It's pretty cool that I recognize that I do both short form and long form.
11:01So let's just say mix, and the system can label it whether it thinks it makes a stronger short form or stronger long form content. When should a fresh deck get generated? I'll just go with the recommended.
11:10Do I want this to happen nightly? Do I want some sort of button in the dashboard that I can click to refresh and bring in new cards? Well, I'll just take the recommended approach.
11:17To make this feel like truly an agentic OS, I'm just gonna have this run nightly on a schedule so that way I come to fresh ideas every morning. What happens when you act on a card? So if I actually film something, what am I going to do?
11:27It recommends that I have two different buttons. One button for if I actually made the video and another button if I wanna skip it and I'm not interested in the topic. So let's go with that.
11:35How much should be on each card beyond the topic, the hook, and the anchor? This is an opportunity for you to customize this to your style. It recommends that I have a lean card with just the topic and the hook, couple of other things, and I'll riff the rest on camera.
11:48That's fine. But I'll do lean and three beats with like three short talking points underneath the hook just so I have a sense of maybe the structure of this video. I always find that to be helpful, and I'll submit these answers.
11:58So this took some time to finish and it built out all of these cards for me from my Obsidian vault, actual coaching calls that I've had, actual problems that we've solved, transformations my members have undergone, all of this stuff. Shows me a mix of short form and long form content, a couple of bullet points outlining the structure of the video, the coaching lesson that's at the core of it.
12:18And I can mark this as something I filmed or I can choose to skip it. And anytime I want, I can just click re deal. It's going to look at the vault.
12:25Going It's to generate brand new topic ideas for me so I never run out of content. One last thing that you can always do to supercharge this is you can tell Fable, you now know everything about this project.
12:36Suggest five ways to improve this agentic OS to be optimized for me. If you wanna learn how to use Cloud Code and grow your income, just check the link in the description. If you wanna see how I built seven $10,000 websites in a single afternoon from one prompt, just check out this video right here.
12:51I'll see you over there.
The Hook
The bait, then the rug-pull.
Most tutorials for a new model show you the spec sheet. This one opens with a live upgrade — a 13-minute session where a single goal-shaped prompt transforms an existing Agentic OS dashboard into a self-refreshing content engine, with Claude making the architecture decisions.
Named ideas worth stealing.
05:20acronym
The GOAL Framework
- G — Ground everything in truth (read before reasoning)
- O — Outcome, not orders (define done in testable terms)
- A — Autonomy over the path (give destination, not route)
- L — Loop in proof (require self-verification)
Four-part prompt structure for giving Claude Fable 5 autonomous, goal-directed tasks that produce verifiable results without micromanagement.
Steal forany Claude Code session where you are handing off a multi-file upgrade or new feature build
How they asked for the click.
VERBAL ASK
12:33link
“If you wanna learn how to use Claude Code and grow your income, just check the link in the description.”
Double CTA — community ( Skool) plus free Notion template for the GOAL prompt. Clean placement at outro, not mid-video.
MENTIONED ON CAMERA
Visual structure at a glance.
