@FinanceYF5: 2/ First change: Give only the goal, not the steps. With previous models, you had to spell out exactly how to do it, or they would go off track. Fable 5 is the opposite — the more freedom you give it, the better it performs. Every step you dictate to it is actually overriding its better judgment with your own.

X AI KOLs Following Models

Summary

Discussing tips for using the AI model Fable 5, advocating for giving only the goal without steps, and granting the model more freedom to achieve better results.

2/ First change: Give only the goal, not the steps. With previous models, you had to spell out exactly how to do it, or they would go off track. Fable 5 is the opposite — the more freedom you give it, the better it performs. Every step you dictate to it is actually overriding its better judgment with your own.
Original Article
View Cached Full Text

Cached at: 07/05/26, 10:31 AM

2/ First change: Give only the goal, not the steps.

With previous models, you had to spell out step-by-step how to do it, or it would easily go off track. Fable 5 flips this around—you give it more freedom, and it performs even better. If you dictate every single step for it, you’re actually overriding its better judgment with your own.

When people ask Matt Shumer the most common question—“How do you actually make those Fable 5 demos?”—he says the secret isn’t a more complex prompt; it’s that the way you prompt is completely reversed.

3/ Just giving the goal can lead to loss of control, so he first sets a few [house rules]—no matter how Fable achieves the goal, these red lines cannot be crossed.

For example, he tells it not to hard-code rules just to handle a specific edge case, but instead to write the behavior into the prompt so it can reason about it. He also has another sub-agent specifically check the output against these rules before it’s pushed.

4/ Second key: Don’t use adjectives like “high quality” as your standard—Fable will settle for a lower bar it has in mind.

He sets a specific, self-checkable hard metric, such as “strangers can’t tell whether this is a render or a real shot.” And the review agent must be a separate Fable in a completely new context—never let the one that made it grade its own work.

5/ After setting the standard, let Fable iterate on that standard in a loop: find the gap, fill the gap, run another round. This can go on for hours or even days. This step is especially useful for creative tasks because there is always something concrete to compare against.

6/ Old work fuels new work: His first 3D forest scene required a lot of effort to write the prompt because there was no reference. But when he later built the Hogwarts demo, he just threw the forest’s code and quality standards at Fable as a reference. Fable could even learn on its own from the old conversation history—which approaches had been tried and failed, and which had worked.

7/ Last point: Clear obstacles in advance—don’t let Fable keep stopping to ask for permission. Give it a budget, give it keys, let it make its own decisions, and only come back to you when it’s truly stuck or when only you can make the call. How do you usually talk to Fable?

Read the original:

That’s all.

If you like this topic:

  1. Follow me (@FinanceYF5)
  2. Like + retweet the first post below

Similar Articles

@FinanceYF5: 2/ He never looks at benchmark numbers when evaluating models. The only thing he truly cares about is: [The shape of the model's thinking] — How deeply can it understand user intent? — How far can it iterate in its thinking? — Does it make you feel like there's someone on the other side? Fable gave him this sense of aliveness. 'It feels like returning to 2023'

X AI KOLs Following

This tweet emphasizes that when evaluating AI models, one should not only look at benchmark numbers but focus on the model's 'shape of thinking' — the depth of understanding user intent, the ability to iterate in thinking, and the feeling of 'someone on the other side'. The author believes Fable excels in this regard, reminiscent of the experience in 2023.

@FinanceYF5: Someone gave Fable 5 a very casual instruction — "write code to make a realistic Eiffel Tower" — and it was done in one go. The result was so good that he couldn't describe it, just saying "it understood". The tower is to real scale, built purely with code to create an open iron frame, and it also has day/night switching and a blinking light at the top. The publisher said that Fable 5's specialty is taking…

X AI KOLs Following

Fable 5, with just a single vague instruction "write code to create a realistic Eiffel Tower", generated a complete 3D model in one pass, including accurate proportions, open ironwork, day/night switching, and lighting effects, demonstrating its ability to understand vague instructions and get it right the first time.

@aiwarts: https://x.com/aiwarts/status/2073962247240773984

X AI KOLs Timeline

Detailed introduction to the most practical prompt engineering tips before Fable5 goes offline, including letting the model run experiments autonomously, optimizing workflows, and reducing token consumption through subagent allocation, along with other practical experiences.

@RealCodedAlpha: After seeing this suspected Anthropic internally optimized Fable 5 Prompt structure, I realized: most people are writing Prompts wrong from the start! Many think a good Prompt = longer, more complex, more techniques, more role settings. But the structure given by this image is very…

X AI KOLs Timeline

This tweet introduces a Fable 5 Prompt structure that is suspected to be internally optimized by Anthropic, emphasizing that writing prompts should focus on goals, boundaries, and acceptance criteria rather than lengthy instructions. Especially for stronger models like Claude and Codex, avoid teaching the model how to think.