Hi AInauts,

Welcome to the latest issue of your favorite newsletter.

This issue is for everyone who felt their pulse jump when Fable 5 came back, then almost pasted the same old monster prompt into it. Please do not.

The smartest model does not need a novel. It needs a clear assignment, clear boundaries, and criteria for what "done" means. You will find all of that in our master prompt.

We also show you which tasks truly deserve Fable, where free AI APIs are enough, and why NotebookLM now turns your own sources into Shorts.

Here is what we have for you today:

Let's go.

Fable 5 is back. If you missed the drama: first we had the hype around Claude Fable as the most powerful AI model, then the reality check when Fable disappeared again and open models suddenly looked like basic equipment.

The lesson back then: do not depend blindly on a single frontier model.

Now Fable is back. And on X, the usual thing is happening: prompt lists, 46-hour demos, cost hacks, agent stacks, and fear of missing out.

For now, Fable is only included in Pro/Max/Team until July 7, and you can use up to 50 percent of your weekly limit on it. After that, it moves to usage credits, meaning pay-as-you-go on top of the subscription.

Anthropic says Fable should return to subscriptions once there is enough capacity. When that happens is still open.

So the practical question is: how do you use Fable sensibly right now without burning your token budget?

We have learned: the more precise the prompt, the better the result. With Fable, that flips.

The most important basic rule is this: Fable 5 needs less classic prompting.

Old detail prompts can even get in the way, Anthropic says. They look diligent. In reality, they often take away the room where Fable is most useful.

Which makes sense when you look at it: you do not want to spoon-feed the smartest model every sentence. You want to give it a clean playing field.

Goal. Boundaries. Sources. Stop rules. Evidence. Let it figure out the rest along the way.

Or put differently: you do not need to prescribe every intermediate step. You mainly need to say how a good result will be recognized.

Fable is expensive, slow, and strong. So its use has to be deliberate, otherwise you are firing a cannon at a tiny problem.

"Make me a marketing strategy" is short and punchy. But short is not automatically good.

A good Fable assignment goes further. It includes context, sources, open decisions, and clear goals with quality criteria. You want to be able to tell whether the work it delivers is actually good.

Every built a useful filter prompt for this: do not let Fable start working immediately. First let it find the right job candidates. Here is our compact everyday version:

The first cost lever sits before the run. It is called task selection.

The OCR trick, where you push context as an image so Fable can read it more cheaply, is neat as long as it still works. For normal work, the better question is simpler: does Fable need to do this at all?

With Fable, the bottleneck is often not the model. It is what you did not say because you were not clear on it yourself yet.

That is why Fable should not start building immediately. It should find the gaps.

Your unknowns come in several kinds: what you said, what you have not decided yet, what you are probably assuming without writing it down, and what you do not even have on your radar.

That is exactly why Fable should first find your blind spots.

If you cannot describe what you want, give references. Old files, screenshots, a good example, a bad example, source code, customer emails.

Anything, as long as it is relevant context.

Fable is good at pulling patterns out of material.

It gets harder when it has to guess your gut feeling.

Loop engineering is the next thing everyone is talking about. For everyday work, it does not need to be as big as it sounds.

At its core, a loop is just this: a small next step, a stop rule, evidence, and a decision for the next step. That is the difference between "Fable, do it" and real delegation.

Fable may run for a long time. But it must not drift endlessly and burn your token budget.

Fable must not become your default model. That is the most important cost tip.

It is the "God mode" for tasks with planning, multiple sources, judgment, long execution, or hard verification.

Tip: inside Claude chats, you can also switch between models.

If Fable itself notices that a part would be cheaper with Sonnet, Opus, or a human, it should say so and not keep calculating out of politeness.

Put this into the assignment:

That creates a practical gain: Fable does not do everything. Fable helps you decide what Fable is worth in the first place.

The new Anthropic Economic Index for June 2026 confirms that 54 percent of Claude Code sessions run on expensive Opus. In chat, it is only 10 percent.

Meaning: autonomy and token use rise together. Better models also like using their peers, even when a simpler, cheaper model would be enough.

The better pattern is: route simple work downward, thinking work upward.

And last but not least: set Fable Effort to high as the default. Use xhigh only for the hardest tasks, and medium/low for routine. Lower effort on Fable often beats xhigh on older models.

Fable 5 is not hungry for your elaborate prompts. It is a very powerful delegation model that needs a clear assignment with boundaries.

If your working mode is sloppy, Fable makes it more expensive. If your working mode is clean, Fable can carry tasks that used to be too large for a chat.

Use it to solve one real knot in your workday: too many tabs, too many notes, a half-described automation, a backlog with no order, a project that has been stuck in your head for weeks.

Write a handoff to Fable that someone on your team would understand.

Here is the Fable 5 Master Prompt that connects all of these tips.

If you have only used AI in the chat window so far, an API is the next lever. It lets you put models into your own workflows: agents, small scripts, tests, content checks, coding helpers.

Here are the most important highlights:

Pick one provider and test it with a real task and several models. The clearer the assignment, the better the results.

Our rule of thumb: free APIs are for prep work. Summaries, sorting, first drafts, log explanations. Anything with customer data, code merges, or legal risk does not belong there.

P.S. Here is a starter prompt so you can use this for your own work.

We have loved Google's NotebookLM from the beginning. After audio, infographics, and video overviews, there are now Short Video Overviews: about 60 seconds, vertical, based on your own sources.

According to Google Help, the format currently runs in English and is first available for paying Google AI Pro/Ultra users, with free availability coming later.

Some people call it "TikTok for PDFs". Learning material, just snackable and scrollable.

We already covered NotebookLM as a learning machine in a German Deep Dive, as a graphics hack, and as part of Google's larger source-content studio.

Language note: the older NotebookLM Deep Dive is available only in German, so we did not link it in this English issue.

The new part is this: the output format now fits Shorts, Reels, and quick community explanations directly. For courses, newsletter recaps, or Skool posts, that is pretty obvious.

We wanted to test it right away with our Fable 5 piece above. Not bad. The prompt is in the video description.

Done. See you in the next issue.

Reto & Fabian from AInauten