**
On June 9th, Anthropic released a new model called Claude Fable 5.
On that very day, I let it handle my entire "automated note selling system."
As a result, daily sales nearly doubled in just over a week.
To be honest, I was just thinking, "A new model is out, might as well try it," so I'm the one most surprised by this.
In this post, I'll write about everything: what I had it do, what was fixed, and even the prompts I used.
By the way, I usually share the latest movements of this system and detailed settings that I can't fit into an article on my LINE, so if you're interested, check the link at the end.
Letting the AI Handle a Successful System on the Day Fable 5 Was Released
First, some context.
Fable 5 is the first model from Anthropic that makes "Mythos-class" capabilities—which they previously withheld from public release due to high performance—available to everyone with safety mechanisms.
It recorded 80.3% on the coding benchmark (SWE-bench Pro), which is a figure that surpasses not only the previous Opus 4.8 but also GPT-5.5 and Gemini 3.1 Pro.
By the way, the "safety mechanism" means that only requests in dangerous areas are automatically switched to older models. According to official data, this doesn't trigger in over 95% of sessions.
In short, for normal business use, you won't feel any difference.
However, what I focused on wasn't the benchmark numbers, but the fact that it can "carry out long and complex tasks without losing context midway."
What does this mean?
It means we can now outsource the "improvement" process itself—reading all the files of an entire system, cross-referencing them with sales data, and judging what needs to be fixed—which previously only humans could do.
With conventional models, if you gave them 10 files, they would often forget the content of the first few and make irrelevant suggestions. Fable 5 almost never does that.
A clear example is a report from overseas where "migrating 50 million lines of code went from 2 months to 1 day." Essentially, it means it can proceed with work while fully understanding a massive whole.
My system has about 40 files in total plus sales data, so this "ability to read through everything" is directly effective.
That's why I decided that "letting it improve an existing system" is the smartest way to use it, rather than just playing with new features.
There's one important perspective here.
A model evolving doesn't mean "the system becomes obsolete."
It means "the speed of system improvement increases."
For those who have a system, every new model release is a tailwind.
The reason I don't do the improvements myself is threefold:
- Doing it yourself introduces "assumptions" (judging why things sell based on your own intuition).
- Outsourcing means handing over the system's contents (leaking the know-how).
- AI can read all data with zero emotion (processing volumes of data overnight that are impossible for humans).
Frankly, improvement is the task of "reading all data and finding the differences," so AI is better suited for it than humans.
My intuition might say, "I feel like empathy-based posts are working this month," but if I let the AI read it, it returns with data-backed findings like, "Posts with high save rates consistently have this element in the intro."
This difference is huge.
The Overview of the Automated Note Selling System
This is the configuration I'm running:
- Concept Design (Consolidating who, what, and how to sell into one md file)
- Automated Post Generation (AI creates 10 posts a day based on the concept)
- Automated Posting (Flowing to Threads/X at set times every day)
- Note Guidance (The flow from Post → Profile → Note article)
- Data Collection (Impressions, saves, and sales are automatically accumulated)
- Automated Improvement Loop (Next posts change based on the accumulated data)
To supplement each part slightly:
The concept md in step 1 is the starting point for everything. Everything from step 2 onwards moves based on the "who, what, and how" written here.
In step 2, the AI creates 10 posts a day on its own after reading the concept md. Humans aren't writing them anymore.
Step 3 is just putting them on an automated posting tool, so I don't even press the post button.
Step 5 is quietly important; "which post was saved," "which post led to a profile click," and "how much was sold that day" are recorded automatically every day.
Step 6 is a loop that reads this data every 3 days and reflects it in the next 30 posts.
The point is that this system isn't "finished once built," but rather the 5 → 6 → 2 improvement cycle keeps spinning.
People often say, "It looks hard to build," but the heavy lifting is only the first time.
You only use your brain when initially building the concept md and evaluation criteria. After that, the more it runs, the more data accumulates, and the more data accumulates, the higher the accuracy of the improvements.
Since it's a structure that gets stronger over time, I honestly think the people who start early win this game.
So what are humans doing? Actually, only three things:
- Deciding the genre and concept at the beginning (this is a human's job).
- Deciding whether to adopt the AI's improvement suggestions (I'll write about this in detail later).
- Occasionally looking at the data and investigating if something seems off.
In terms of daily work time, it's less than 10 minutes.
And what I had Fable 5 do this time was the most upstream task of the cycle: "Cross-reference all files of this system with recent sales data to identify areas that need fixing."
I didn't do anything difficult for the data transfer; I just let it read the set of system files and the data containing sales, impressions, and saves as they were.
Many people think, "I have to organize it before handing it over," but it's the opposite.
If you hand over raw data, the AI will organize it on its own and pick up correlations that humans don't notice.
The prompt I threw was basically this (I'll leave it here so you can copy and paste it):
"Read all files of this system and compare them with the sales data to identify all the differences between 'when it sells' and 'when it doesn't.' Give me only the data differences, not impressions. Next, propose which file and which judgment criteria of the system those differences should be reflected in, as a set of filename and modification proposal."
The point is that I restricted it to "only data differences, not impressions."
If you don't include this, the AI will mix in general advice like "Increase posting frequency," the kind of thing written everywhere.
By limiting it to data differences, only improvement points unique to my system come out.
After the differences are out, the prompt for actually making it fix them is this:
"I will specify which of the proposals I will adopt, so rewrite the corresponding files directly. Show me a list of the differences before and after the rewrite at the end. Do not change a single character in the parts I haven't specified."
"Do not change a single character in the parts I haven't specified" is seriously important. If you don't include this, the AI will "improve" other areas out of goodwill. When letting it touch a working system, the ironclad rule is to restrict the scope of changes.
Three Differences Fable 5 Found Between "When It Doesn't Sell"
So, there were three differences that Fable 5 brought up.
The moment I saw the first one, I thought, "This will definitely change sales."
Difference ①: Selling notes had "the sentence right before buying"
This is the conclusion Fable 5 reached after comparing all selling and non-selling notes.
The concepts of selling notes had the "emotions that come to the reader's mind right before buying" verbalized and included.
The ones that didn't sell were only designed up to the "target" and the "problem."
Specifically, the concept of a non-selling note was at this level of granularity:
"30s female · wants to get back with ex · teach her how."
And the concept of a selling note was like this:
"30s female · wants to get back with ex · buys the moment she thinks, 'If I do exactly what this person says, I might not have to look at my phone at midnight and feel anxious anymore.'"
Do you see the difference?
The former ends with "what to sell."
The latter is designed down to the thoughts in the head 0.5 seconds before opening the wallet.
This was the dividing line for sales.
So, I had it add "verbalization of emotions right before purchase" as a mandatory item in the concept design md.
Just adding this one item changes everything. Since the concept is upstream, the word choice of the generated posts, the title of the note, and the way the body text hits—everything starts being written toward that "thought right before buying."
A modification of one item in one file ripples down to all downstream products. This is the beauty of systematization.
By the way, I also had Fable 5 output the knack for writing this item: the criteria is "writing the true feelings the reader can't tell anyone, in the reader's own words."
"I want to resolve my anxiety" is weak.
If you can write, "I'm so afraid of being ghosted again that I can't press the send button," it's a pass.
It's the same in other genres. For example, in the side-hustle niche, instead of "for people who want to earn from a side hustle," use:
"People who look at their balance a week before payday and put one item back in the basket at the convenience store while sighing."
If you can write that far, it becomes a note that looks like an ad visible only to that person.
Difference ②: Post evaluation criteria were too focused on impressions
The evaluation criteria for the automated improvement loop were "impression-heavy."
However, what correlated with sales wasn't impressions, but "number of saves" and "profile click rate."
This means posts that get impressions and posts that open wallets are different things.
Anyone doing list marketing would understand this, but the number of leads and the number of conversions don't necessarily correlate.
Posts that get impressions are the "empathy, relatable, humor" types.
But for posts that open wallets, movements like "I want to save this and read it later" or "Let's check this person's profile" are happening.
Going viral feels good, but the numbers directly linked to sales were the more modest ones.
By the way, these two numbers can be seen by anyone for free on X Analytics.
Just by reviewing your posts in order of "saves" instead of "impressions," you should find the patterns that sell.
For reference, saved posts had three things in common:
- They contained specific steps that can be used later (doesn't end the moment it's read).
- They contained numbers or proper nouns, making people want to keep them as a "memo."
- There was room to apply it to one's own situation (in a form that can be used as-is).
Conversely, "good stories" aren't saved. Impressions are consumed on the spot and end there.
When I had it rewrite the evaluation criteria to "focus on saves and profile clicks," impressions dropped slightly, but sales went up—an interesting movement.
Just by changing one line of the evaluation criteria in the automated improvement loop, the direction of all generated posts changes, so people who have a system should review this first.
Difference ③: The order of the body text was "solution too early"
Notes that didn't sell gave away the solution too much at the beginning.
If you show the answer before the reader realizes "this is my problem," they finish reading before they feel the value.
Selling notes used "30% of the total for verbalizing the problem."
In a 7,000-character note, the first 2,000 characters are used not to give a solution, but to "describe the reader's situation more accurately than the reader themselves."
Only after the reader is in a state of "Why do you know me so well?" do you provide the solution for the first time.
In this order, even the same content is read as something "valuable."
And for those wondering "how to write the verbalization of the problem," I'll leave the procedure I actually use.
I have the AI write out 10 monologues that the target is thinking in bed at night.
Things like "It's been a week and still no read receipt," "I can't talk to my friends about this," or "My search history is all about getting back with an ex."
Just by rearranging those 10 and placing them at the beginning of the body text, the problem verbalization part is almost complete.
A common mistake here is writing the problem in "generalities."
Phrases like "Getting back together is hard, isn't it?" or "It makes you anxious, doesn't it?" that apply to anyone won't hit anyone. Only when you specify it to the level of a monologue does the "This is about me" moment occur.
This was also something Fable 5 brought up as a data difference. It's a basic of sales when you think about it, but it was a part that hadn't been incorporated into the automated generation system.
Things humans would do naturally are missing from the system design.
Finding these "holes that were known but not implemented" is the best part of letting AI do the improvements.
Design Philosophy When Leaving Improvements to AI
There are three principles I learned from trying this for leaving improvements to AI.
Principle ①: Improvements to AI, adoption decisions to yourself
There were 7 proposals from Fable 5 in total.
I adopted 3.
Among the 4 I rejected was, for example, a proposal to "increase the number of posts from 10 to 15 a day."
The logic was sound, but I had already verified in the past that for my reader base, the timeline pressure would be too strong and they would end up muting me, so I rejected it.
If you swallow everything whole, the system will lean toward AI generalities and, conversely, stop selling.
In fact, I once adopted all proposals in the past, and the numbers dropped that week.
Even if each one seems correct, when combined, it becomes an "account found anywhere."
AI is a genius at finding differences, but the materials for judging "whether to adopt that difference"—past verification history and the feel of the readers—exist only within yourself.
Principle ②: Always reflect improvements in the "files"
Don't just say "do this from now on" in the chat and end it; have the AI rewrite the system files themselves.
The reason for doing this is to turn improvements into assets.
Instructions in a chat disappear, but judgment criteria placed in a file continue to be effective for everything from the next generation onwards.
In my case, I divide the rewrite destinations by role.
"Who to sell what to" is the concept file, "which post to judge as having grown" is the evaluation criteria file, and "in what order to write the note" is the structure file.
Since I can see which file grew with each improvement, it feels like the entire system is becoming smarter and smarter.
Principle ③: Fix from upstream (Concept > Post > Wording)
When it comes to improvements, you tend to want to fix from the downstream, like the wording of a post.
However, the order of effectiveness for sales is "Concept > Evaluation Criteria > Body Structure > Wording."
The main reason sales doubled this time was undoubtedly the concept part of Difference ①.
The reason is simple: one upstream location affects everything generated from there, but one downstream location only affects that one piece.
If you fix one line of the concept, all future posts and notes will change. If you fix one ending of a post, only that post changes.
If the workload is the same, it's better to use it where it's effective.
No matter how much you polish the downstream, if the upstream design doesn't sell, the numbers won't move.
People on the System Side Profit Every Time AI Evolves
So, what's happening now?
I reflected the improvements in mid-June, and from there, daily sales have stabilized at nearly double.
The most effective part was undoubtedly the concept part of Difference ①. Immediately after reflecting it, the flow of "Save → Profile → Purchase" from the generated content changed visibly.
Work time remains unchanged, almost zero.
This incident made me certain of one thing:
The people who profit most every time AI evolves are "those who have a system."
Fable 5 itself can be used by anyone.
But people who don't have a "target for Fable 5 to improve" can only try it out when a new model comes out.
There are people who try a new model every time it's released, say "it's amazing," post their impressions, and end there.
And then there are people who, every time a new model comes out, let it read their system, improve it, and raise their sales figures by one level.
Even though they're using the same AI, one is consumption and the other is investment.
Because I had a system and data, my sales moved on the day the model evolved.
Models will continue to evolve.
In six months, a model smarter than Fable 5 will be out, and I'll do the same thing again.
This loop itself is an asset.
Every time that happens, I believe the gap between those who have a system and those who don't will widen at an incredible pace.
I'm sometimes asked, "Isn't it too late to start now?" but it's the opposite.
Latecomers can build a system with a smart model from the start, so they can skip the parts I struggled with in the past.
When I started building this system, I had to manually fix the AI's output every time, but people building it now will hardly need that process.
Whether it's late or not is decided not by the time of entry, but only by whether you start building a system.
Now, for those who read this far and thought, "I want to move to the system side."
Actually, an incredible project is underway right now.
I'm doing a real-time challenge called "Making a non-tech-savvy housewife in her 40s earn 200,000 yen using only Claude Code automation, with a 30-minute-a-day limit."
While you're sleeping, the AI finishes writing a note, and when you wake up in the morning, it's selling on its own.
I'm showing everything, including failures, of how "leaving it all to AI and having it earn for you" is becoming a reality.
The steps to build a system from scratch, templates for concept design, and the movements of people who are actually earning are all being shared here, so honestly, I think you're missing out if you don't see it.
What you can receive by participating ↓
✅ The entire process of a non-tech-savvy housewife in her 40s achieving 0 to 200,000 yen (disclosed at a level you can imitate as-is)
✅ Distribution of the contents of the Claude Code automation actually used in the challenge via commentary columns
✅ The latest know-how on AI x Note sales (over 50 people reporting results; a case where someone who was a know-how collector made 147,000 yen in one month)
Participate here (completely free, there's a limit on the number of people, so please join early)
https://line.me/ti/g2/2NjuIznaLxS8gyB0eKLdMOQxQvpcYUYj1e9TlQ
