How to Extract Everything from Your Head While Claude Fable 5 is Still Available (Until July 7th)

The period during which you can use Claude Fable 5 without additional charges is currently announced by Anthropic as being "until July 7th."

This schedule may change in the future, but in any case, the fact that the current Fable 5 is "time-limited" remains the same.

Even when the period ends, Fable 5 itself won't disappear, but it will be removed from the standard subscription quota, and continuing to use it will require additional payment.

Now, there are many articles circulating on the timeline about "things you should make Fable 5 do before it ends." I also wrote an article in that vein yesterday.

But today, I'm going to talk about the opposite side.

This is about how not to be in trouble "after Fable 5 is gone."

If yesterday's article was about "using it to the fullest," today's is about "leaving things behind." The content is completely different, but please think of them as a set.

To jump to the conclusion: the more you have delegated work to Fable 5 since its revival and been impressed by how "obviously smart" it is, the higher the possibility you will be in trouble after it ends.

The reason is simple: when you work with an excellent AI, your instructions become increasingly sloppy.

You can get away with saying, "Hey, do this nicely."

Once you've explained something, it remembers and uses it without you having to say it again.

It senses and compensates for conditions you forgot to mention.

It's comfortable, isn't it? I've been in this state since yesterday too.

However, the one who holds the substance of that "nicely" in words is Fable 5, not you.

To be precise, the conversation logs themselves will remain without being deleted even after the period ends. What you will lose is the "reader" who can accurately grasp your intentions from those cluttered logs and sloppy instructions.

To use an analogy, it's like being left with only the scribbled notes of a departing secretary. The notes exist. But only that person can read them fluently.

I realized this yesterday while I was checking my work folders with the AI.

The production procedure for delivery-use e-learning videos (commentary videos for training) that I had just established with the AI at the end of June. No matter how much I searched for files or notes, they were nowhere to be found. They only existed within the past conversation logs. When I opened the work folder, it was empty.

In other words, I had just "put it in the chat and finished." If I hadn't noticed, the procedure would have remained buried in the conversation, never to be dug up again.

In this article, I will summarize four "handovers" you must absolutely do while the excellent AI is still here.

• Recover deliverables that only exist in the chat.
• Have it create a handover document for things it was "sensing" for you.
• Rehearse the handover (this can only be done now).
• Implement two rules so that future work is automatically preserved.

No difficult technology is used. No programming is required. All you need is the "way to ask."

Chapter 1: What Disappears and What Can Be Saved

First, let's organize our thoughts. Fable 5's intelligence can be divided into two types.

One is "situational intelligence." Answering questions accurately, reading long documents correctly, and completing complex tasks.

This is the performance of the model itself, so once the period ends, that performance leaves your hands. There's nothing you can do about this.

The other is "lasting intelligence." The way of proceeding with work established during your conversations. Policies decided after many exchanges.

Completed texts and systems. The "template" of "if I ask like this, it returns like that."

This is not the model's performance, but "information."

If it's information, it can be made into a file. If it's a file, it can be passed to the next AI.

What you need to do is separate these two and convert the "lasting intelligence" into files during the period. That's all.

The Better the AI, the More It "Senses." That Becomes a Trap

Let's talk about another important structural point.

Models of the Fable 5 class read between the lines of instructions.

For example, the reason a correct invoice comes out with just the phrase "make last month's invoice" is that the AI senses and supplements the background, such as hourly rates and recipients, from the flow of conversation and past files.

In a human workplace, it's like working with a super-competent secretary. "Do the usual thing" works for everything.

The problem is that this secretary has a fixed resignation date.

Moreover, there isn't a single handover document yet.

The next AI that comes can also read the past interactions themselves.

However, it won't grasp your "usual thing" with the same precision as the current secretary.

That's when you realize: the reason those sloppy instructions worked was because the model was smart, and you hadn't been able to explain your own way of working in words.

The important thing here is that I'm not saying the next AI you use will be incompetent.

Sonnet and Opus have sufficient ability to execute explicit procedures.

However, to be honest, the difference isn't just in the "ability to sense intentions from vague instructions." The quality of the output itself is also different. You can't hide that.

But how much that gap widens changes significantly depending on whether you have explicit procedures and samples at hand.

If the model changes while instructions remain sloppy, the misalignment of intention and the drop in quality will come at the same time. If procedures and samples remain, at least the template of the work won't collapse.

So the countermeasure is simple: just put into words the parts that were being sensed for you.

And the one who can do that verbalization best is Fable 5 itself, which knows all your context right now and can accurately pick out the key points from long conversations.

In fact, the task of "reading cluttered long conversations and organizing them correctly" is exactly where the difference in model intelligence shows the most.

In other words, the extraction work itself is the job you should make Fable 5 do. Even if you try to make the next AI do this, it won't be able to pick things up with the same quality.

Chapter 2: Must-Do ①. Recover Deliverables That Only Exist in Chat

The most immediate effect comes from recovering from past conversations.

The Story of How I Almost "Lost" Something

At the end of June, I was verifying "how to make e-learning videos" with the AI.

I produce commentary videos for training as my job, and I was working on which processes to leave to the AI and how to stabilize quality.

I got to the point of solidifying the division of labor and procedures.

That conclusion existed only within the conversation.

Yesterday, while checking folders for another task, I noticed it. I was planning to mass-produce videos using this procedure in July, but the procedure manual was nowhere to be found.

If I had forgotten the conversation entirely, I would have had to start the verification from zero.

Recovery was easy. I just had Fable 5 read the past conversation and summarize it into a single file as a procedure manual. It took a few minutes.

You definitely have these kinds of conversations too.

• Texts that were finally completed after many exchanges
• Things decided as "let's go with this policy from now on"
• Procedures that worked well after trial and error
• The overall picture of your work or business that the AI organized for you

These remain if you open that conversation. But conversations flow away. A month from now, will you be able to remember which conversation it was?

How to Do It

Roughly recall the work you did with the AI in the past 1-2 months and list them in order of "how much trouble I'd be in if it disappeared." Then, open that conversation and ask like this:

1This conversation contains important conclusions that I want to continue using.
2Please re-read the entire conversation and summarize the following into one file (Markdown format).
3
4- The final version of the completed deliverables (text, structure, code, etc.)
5- The decided policy and the reason for it
6- The established procedure (at a level of detail that can be reproduced by another person or another AI)
7- Points of caution, failures, and countermeasures
8- Things that remain unresolved
9
10Rules:
11- Do not supplement or create content that does not exist in the conversation
12- Do not summarize too much. Leave the procedures and deliverables in full without omitting anything
13- Write in one line at the beginning "which job this file should be referred to for"

Save what comes out.

At this stage, the storage location doesn't matter.

The priority is that it exists as a file. (By the way, I put mine in a memo app called Obsidian. I'll talk about the reason in Chapter 5.)

The trick is not to aim for perfection.

You don't need to recover every single conversation.

Just pick up as much as you can, starting from "things that would take hours to recreate if they disappeared."

Even recovering just three will save you considerably.

Chapter 3: Must-Do ②. Have It Create a Handover Document for "Things It Was Sensing"

The chapter I want to convey most in this article is this one.

If you are working comfortably with Fable 5, a large number of "unwritten rules" have accumulated behind the scenes.

Your stylistic quirks. The boundary for "don't do this." The passing line for deliverables. The location of frequently used materials.

If you try to write these down yourself, you will almost certainly fail.

Because they are so obvious to you, you can't verbalize them.

However, you are an object of observation from the AI's perspective.

Because it's looking from the outside, it can verbalize them. So, have Fable 5 itself write it.

In the conversation you've been interacting in the longest, please paste this:

1You will soon be leaving this job.
2Please create a handover document for the successor AI (assuming it has weaker "sensing ability" than you).
3
4Based on our conversations so far, please write down what you have learned about me in the following categories:
5
6- The overall picture of my work (what kind of person I am)
7- My preferences (writing style, tone, format of deliverables)
8- Instructions I often give and their true meaning
9  (Example: When I say "do it nicely," what does that actually refer to?)
10- Things that must absolutely not be done
11- Criteria for passing deliverables (what kind of situation makes me say "OK")
12- Things I often forget to say and how to supplement them
13
14Rules:
15- Write only what could be confirmed in actual conversations. Do not fill in with speculation
16- For items you are not certain about, clearly state "unconfirmed"
17- Write with enough detail so that the successor AI can start working with the same quality from tomorrow just by reading this file

Be sure to read and check the handover document that comes out yourself.

The AI's observation might be wrong.

But that's fine.

The act of correcting it by saying "this part is different" itself becomes the work of determining the rules of your job.

One more thing. If there is work you repeat every week or every month, ask in that conversation: "Make the instructions for this task into a prompt that can be used by copy-pasting so that another AI can reproduce it with the same quality. Attach one of your actual outputs this time as a sample."

The point is to "have it attach a sample."

With just a procedure manual, the result will fluctuate, but with a procedure manual with a sample, even a model with weak sensing ability will align itself with the sample.

Chapter 4: Must-Do ③. Do a "Rehearsal" of the Handover

This is something that can only be done during this period.

You won't know if the handover documents and instructions you made in Chapters 2 and 3 actually function until you actually try running them with another model.

And now, you can do this.

1. Paste the instructions you made into Sonnet or Opus and have them execute it.
2. Look at the result and note down where it deviated or where you were asked questions.
3. Take that deviation back to Fable 5 and have it fix the instructions, saying "this part didn't get across."

This third step is important.

After the period ends, even if the instructions don't work well, the AI that can fix them won't be at hand (unless you pay extra).

The back-and-forth of "having the smart AI make the handover document, testing it with the weaker AI, and having the smart AI fix it" can only be done now when both are at hand.

You don't need to do it for all instructions. Just choose one job you use most often and do one round-trip; the precision of the handover will change completely.

As a bonus, this comparison produces another result.

You will know based on actual results the line of "Sonnet is enough for this job" and "I'll pay extra for Fable 5 for this." You can use this directly for your payment decision after the period ends.

Chapter 5: Two Rules to Automatically Preserve Future Work

Finally, here is a system so that future work doesn't end with "just putting it in the chat." There are only two rules.

Rule 1: Say "Make the current conclusion into a file and save it" at the end of the conversation.

In environments where the AI can directly create files like Claude Code or Cowork, decide on a save folder and have it write there.

In a normal chat, have it output in Markdown format and save it yourself. With just this, deliverables will structurally stop flowing away with the conversation.

Rule 2: Once the files increase, have it create one table of contents.

Place a table of contents file at the entrance of the folder that lists one line for each file, what is written in it, and when it should be read.

To make it, just ask the AI: "Create an index for this folder. Clearly state at the top the file that an AI coming for the first time should read first."

The reason the table of contents is important is that if you make the AI read all the files every time, it's slow, and irrelevant information gets mixed in, lowering precision.

Make it so you can just give the table of contents and say "read only the necessary parts." Even if a new AI comes, it can grasp the whole picture in a few minutes from this one sheet.

My Case. I Had Fable 5 "Inspect" the Obsidian I've Operated for a Year

By the way, I have been managing all my work information in a free memo app called Obsidian for about a year.

Account designs, product materials, past manuscripts, daily logs. I've accumulated everything here.

And I've always worked by having multiple AIs like Claude Code, Cursor, and Codex read this Obsidian.

If I ask for a draft of an article, it automatically refers to the templates and stylistic rules of past manuscripts.

If I ask for a product description, it reads the product materials before writing.

Even when I switched models, the work kept moving not because the AI was smart, but because this "knowledge that AI can read" was at hand.

I can say the effect of the system written in this chapter as a year's worth of actual experience.

So, what did I make Fable 5 do first when it was revived yesterday?

"Inspection."

I had it read the entire Obsidian and past work to scrutinize what was missing, what had become old, and what was scattered.

To give the evaluation first, Fable 5's judgment was: "This system is 90% complete. Almost no one has built it to this extent."

It seems that what I've been steadily accumulating for a year had properly taken shape.

But the important part starts here.

Even from that "90% complete" Obsidian, omissions were found.

That is the "production procedure that remained only in the conversation" I wrote about in the introduction.

Besides that, several holes in the index and materials that were about to become double-managed came to light.

Even if it's 90% done, things leak.

I believe this is exactly where to use a smart model.

The job of looking across a large number of files and long conversation logs to find "what is missing" shows the difference in model intelligence as it is.

For those who already have a system, use it for inspection; for those who are about to create one, use it for the initial foundation while Fable 5 is here.

The fact that the quality of knowledge determines the precision of AI output hasn't changed since the time ChatGPT first came out.

That's why in the past, everyone used to paste and explain their own knowledge and premises in the chat every time.

Obsidian changed that.

Instead of giving knowledge every time, you store it and have the AI come to read it. I have been nurturing that "storage place" for a year.

This System Does Not Depend on a Specific AI

And everything created up to this point is just plain text files.

For AIs that can directly read files on your computer like Claude Code, Cursor, or Codex, they will start working just by you saying "read this table of contents and handover document first."

Even for AIs that cannot directly read folders on your computer like ChatGPT or the browser version of Claude chat, it will work the same way if you attach or paste the files. (Note that the Claude desktop app also has a function to directly read and write files on the computer.)

It's not in a format that can only be opened with a specific app.

This is the important part.

The work started as a countermeasure for Fable 5 going away results in creating a state of "it's okay no matter which AI comes."

It means you can make this the last time you are tossed around by the appearance and end of models.

For People Who Don't Have Time. Three Things to Do in 30 Minutes Today

If you don't have time to do everything, just do these.

1. Recall 3 "conversations you'd be in trouble if they disappeared" (just write them down in a memo on your phone)
2. Paste the prompt from Chapter 2 into the one you'd be most in trouble with and recover it (first, create a state of having "recovered one")
3. Have it create a handover document (Chapter 3. There is a density that only the current Fable 5, which has accumulated the deepest context with you, can write)

Conversely, there are things you don't need to do. Recovering all conversations (there's no end to it, so just "troublesome ones"), clean folder design (you can fix it later), challenging new automations (if you spread yourself too thin during the period, it will be half-baked).

Note that this extraction work itself consumes less of the usage quota compared to heavy production work. This is because the work is centered on reading and summarizing.

Please sandwich it between production-type jobs.

Summary. AI Intelligence is Borrowed. Only What You Make into Files Remains

AI intelligence has "situational intelligence" and "lasting intelligence."

Situational intelligence is something borrowed that you return when the period ends. But lasting intelligence becomes yours if you make it into a file.

• Recover deliverables that only exist in the chat.
• Make things it was sensing for you into a handover document.
• Do a rehearsal of the handover while both AIs are here.
• Leave future work behind with the two rules of "save" and "table of contents."

In yesterday's article, I wrote "let's complete heavy jobs by the deadline."

Today is the continuation of that, about carrying over the things you completed and the "way of working" born until completion to the period after.

Using it to the fullest and leaving things behind. I think this time-limited offer will truly pay off only when these two are combined.

I will also continue the handover in parallel while making Fable 5 work until the end.

For your future self, please start with 30 minutes today.

Finally, a proposal rather than a request.

This article will also disappear once it flows down the timeline.

If you thought "I'll do it later," please save it before it disappears.

The prompts I distributed can be used with any AI even after the deadline has passed.

And if you thought it was helpful, I'd be happy if you let me know with a like or a repost.

There will surely be someone on your timeline who says "oh no" on July 8th. If it reaches that person, this article will have fulfilled its role.

Main Official References

Revival of Fable 5, target plans, and provision period

https://www.anthropic.com/news/redeploying-fable-5

Promotional usage conditions for Fable 5

https://support.claude.com/en/articles/15424964-claude-fable-5-promotional-access