The most valuable thing Claude Fable 5 has is not its intelligence. It is its understanding of you. And that is the one part that expires.
Anthropic has announced that Fable 5 access inside paid plans moves to metered usage credits. Every conversation I have seen about it has been a conversation about pricing. That is the wrong thing to be looking at.
You are not about to lose capability. Opus is an excellent model and it will do your work. What you are about to lose is the months of accumulated context Fable 5 has built up about your business: your clients, your formats, your tone, the seventeen small corrections you made until it finally got your weekly report right.
That context lives in a session. Sessions end. So before the window closes, spend one evening moving that understanding out of the model and onto your disk, where it belongs.
The Idea in One Line
Fable 5 is the best model available at understanding how you work. So before it goes pay per token, make it write that understanding down.
Files survive the deadline. The model's context does not.
That is the whole thesis. What follows is three prompts, run in order, in a single evening. The first one finds your repeated work. The second one teaches any model to do that work your way. The third one makes sure every future session starts already knowing who you are.
What You Actually Lose If You Skip This
Put the two futures side by side. The difference is not the model. It is whether the understanding was written down before the model that had it went away.
The deadline hits, the context is gone
- Opus starts from zero on your work
- You re-explain your business every chat
- Fable's understanding expires with the window
The files survive, nothing is lost
- Opus reads the files and picks up mid-stride
- Every session starts pre-briefed
- Fable's understanding is on disk, permanently
Step One: The Audit Prompt
Open Claude Code in your main work folder. Not a scratch directory, not a demo repo. It has to be the folder with your real projects in it, so the model can see how you actually work rather than how you would describe yourself working. Then paste this.
Study this folder. List every task I repeat every week, ranked by hours spent. Then list the prompts and files you'd need to automate each one. Don't guess. Base everything on what's actually in these files, and ask me up to 5 questions where you're unsure.
You will get back a ranked list of your repeated work. The weekly report. The outreach emails. The invoices. Whatever your version of that is. It will be uncomfortably accurate, because it is reading your files rather than your self-image.
Two clauses in that prompt are doing real work, and it is worth understanding why before you paste it.
- 01"Don't guess." Without this, a model will happily produce a plausible list of tasks a person like you probably does. Plausible is worthless here. You want the list that is evidenced by what is on disk.
- 02"Ask me up to 5 questions." This converts the model from a guesser into an interviewer. The cap matters, because without a number it either asks nothing or asks forever.
That ranked list is the raw material for everything below. Do not skip ahead.
Step Two: The Skill-File Conversion
The audit is not the point. The handoff is. Now you turn each repeated task into a skill file: a short markdown file that teaches any model to do that one job your way.
For each repeated task you found, create a skill file in a /skills folder. Each file should contain: - when to use it (one line) - the exact steps, in order - one full example of a good final output, taken from my real files - the mistakes to avoid, based on what you saw me correct Write them so a model that has never seen my work could follow them. Save each file, then show me the list.
Four ingredients, and they are not equally weighted:
- ✓When to use it, in one line. This is the trigger. A model scanning a folder of skill files needs to know in one glance which one applies.
- ✓The exact steps, in order. Not principles. Steps. The thing you would tell a competent new hire on day one.
- ✓One full example of a good final output. Taken from your real files, not invented. This is the ingredient that carries the weight.
- ✓The mistakes to avoid. Drawn from the corrections you actually made. Your edit history is a specification, and nobody ever writes it down.
The example is the part doing the heavy lifting. Models match a good example far more reliably than they follow a description. You can spend a paragraph describing your tone and get something adjacent to it, or you can paste one thing you actually shipped and get something that sounds like you. Every hour you consider spending on the description, spend on picking a better example.
A description of your work is an opinion. An example of your work is a specification.
This step is the bulk of the evening. That is fine. It is also the only part that compounds. Each skill file you write is a task you never fully explain again, to any model, forever.
Step Three: The Instructions File
Now the brain transplant. One file that carries everything Fable learned about you, loaded by every future session, no matter which model is running.
Write an INSTRUCTIONS.md file (or add to my CLAUDE.md) that teaches a new model how I work: - my stack, tools, clients, and formats - my tone and writing rules, with a short sample - the weekly tasks and which skill file covers each one - what a good day's output looks like, concretely Keep it under 2 pages. Be specific enough that a cheaper model reading it cold gets my work right on the first try.
In Claude Code, name it CLAUDE.md and put it at the project root. It loads automatically at the start of every session. You do not paste it, you do not reference it, you do not remember to attach it. It is simply there, in the model's context, before you have typed a word.
That is the handoff. The next model starts with everything instead of nothing.
Note the two-page cap. It is not an arbitrary constraint. A long instructions file is almost always a vague one, padded with things that sound like context but do not change any decision the model makes. Two pages forces you to answer the only question that matters: what would a competent stranger need to know to get my work right on the first try?
And note the last bullet: what a good day's output looks like, concretely. Same principle as the skill files. Show the model the finish line rather than describing the direction of it.
Do It Tonight, In This Order
The order is not a suggestion. Each step consumes the output of the one before it. The audit produces the task list that the skill files are built from; the skill files produce the inventory that the instructions file points at.
If you only have time for one thing, do the instructions file. It carries the most understanding per minute of setup. The skill files make the work precise; the instructions file makes every future session start informed. On a night where you have twenty minutes rather than three hours, that is the trade to make.
The Honest Notes
Two things I want to be straight about, because a playbook that oversells itself is worse than no playbook.
On the deadline. The free-window timing comes from Anthropic's announcement that Fable 5 access inside paid plans moves to metered usage credits. I am not going to quote you a date, because plans differ. Check your own plan for the exact cutoff rather than trusting a number from a blog post.
On the outcome. A cheaper model with your files gets surprisingly close. It does not get identical. The gap shows up most on genuinely hard, novel problems. Those are the ones where you actually wanted the frontier model's reasoning, and no amount of documentation substitutes for it. And the gap barely shows at all on the repeated weekly work these files cover.
Which is precisely the argument for doing this. The work that a documented cheaper model handles well is the work that eats your week. The work where the gap is real is the work you were going to think hard about anyway. Spend the evening, and you spend the rest of your credits on the problems that deserve them.
Files survive the deadline. Go write them down.
Want this kind of agentic setup for your own brand or product?
I help founders and brands build generative AI systems that actually ship: content engines, agentic workflows, and AI-built sites that are production-grade down to the schema layer. If you want a second pair of eyes on your setup, book a session.
Book a Growth Chat
The three prompts in this post are reproduced exactly as I run them, and are meant to be copied verbatim. Anthropic's own announcement is the source for the premise that Fable 5 access inside paid plans moves to metered usage credits; the exact cutoff varies by plan, so confirm yours before relying on a date. I have written previously about what Fable 5 did with a night of unsupervised production access, which is the same argument from the other direction: context, front-loaded, is what makes an agent useful.