The CLAUDE.md trick that outlives Fable 5

I stumbled on a post this week that made me stop scrolling and actually try the thing before I finished reading it. That almost never happens. The original poster shared a tiny setup trick for Claude that solves a problem I’ve complained about for months: your best AI model only sticks around for so long, and then you’re stuck re-teaching a lesser one everything from scratch.

The mind behind this figured out a clever workaround. And the best part? It costs nothing and takes about ten minutes.

The problem this creator spotted

Powerful models come and go. You get a short window with the smartest one, then it rotates out and you’re back to explaining your tone, your formats, and your rules all over again. Frustrating doesn’t cover it.

Here’s the insight the author landed on. Claude reads a file called CLAUDE.md before every single prompt. Whatever’s written in that file, every future session follows, no matter which model is running.

So the trick is simple: you make the smartest model write that file while you still have access to it. Then its thinking never really leaves. Every session after the model rotates out keeps following the instructions the best version wrote for you.

Why it matters: you’re basically freezing the judgment of a top-tier model into a text file that outlives it. The model changes. Your setup doesn’t.

The step-by-step the original poster laid out

This is where the process gets really practical. Each step has a reason behind it, so I’ll add those as we go.

1. Open the Claude desktop app and go into Cowork. This is the workspace where Claude can actually see your folders and files.
2. Open the folder you work in most. Pick the messy one: reports, client work, content. The messier it is, the more the file has to learn from.
3. Copy the prompt below, paste it into the chat, and send it. No editing needed.
4. Let Claude read the folder. It scans your files, your naming, and how you organize things, so it can infer your habits without you spelling them out.
5. Answer its interview questions honestly. It asks up to 8, one at a time, about how you like to work. Honest answers make a sharper file.
6. Watch it write up to three files: one at the computer level, one at the project level, and one at the folder level. Each covers a different scope.
7. Review each file before it saves. Cut anything that isn’t really you. This is your quality gate, so don’t skip it.
8. Save and close the chat. Done.

After that, every future session, on any model, follows what the smartest version wrote. That’s the whole payoff.

The exact prompt to copy and paste

This talented creator shared the full prompt, and I’m reproducing it word for word so you can grab it as-is:

I want you to write my CLAUDE.md files, instructions every future session will follow.

First, look through this folder: the files, the subfolders, how everything is named and organized, and any existing CLAUDE.md files.

Then interview me: ask up to 8 questions, one at a time, about how I like to work, what I keep having to correct or repeat, what my deliverables need to look like, and anything you can’t figure out from my files.

Then write up to three files:

• A CLAUDE.md in my main Cowork folder: only things true for everything I make, my tone, my formats, my rules for how you work with me.
• A CLAUDE.md in this project’s folder: only things specific to this project, what it is, who it’s for, where things live, what the finished output looks like.
• If any subfolder genuinely works differently (say, one is client-facing copy), a short CLAUDE.md just for that folder.

Rules for writing them: be specific, never vague (“start every draft with the hook, not background”, not “write engagingly”).

Keep each file well under 200 lines, shorter is better. Include only instructions I’d otherwise have to repeat in every chat, nothing you can see from my files anyway. Never repeat a rule across two files. Where a rule isn’t obvious, add its one-line reason. Every line has to earn its place.

Show me each file before saving it, and tell me what you deliberately left out and why.

Why this setup is smarter than it looks

A few things jumped out at me when I read the prompt closely. The expert built in guardrails that most people would never think to add.

• Specificity over fluff: the prompt forces concrete rules like “start every draft with the hook” instead of vague junk like “write engagingly.” Vague instructions produce vague output.
• No duplication: a rule can’t repeat across two files. That keeps the whole system lean and easy to maintain later.
• One-line reasons: when a rule isn’t obvious, the file explains why. Future-you (and future models) understand the intent, not just the command.
• Human review built in: nothing saves until you approve it, and Claude tells you what it left out and why. You stay in control of your own setup.

I love that last touch. It treats the file like a living document you edit, not a black box you blindly trust.

A couple of practical tips

If you want to get the most out of this, here’s what I’d keep in mind based on how the author framed it.

• Pick your genuinely busiest folder, not a tidy demo one. The file learns from your real mess, so give it real mess.
• Don’t rush the interview. Those 8 questions are where you encode the corrections you’re tired of repeating.
• Be ruthless in the review step. Every line should earn its place. Cut anything generic.

The person who posted it also hinted that these files are only step one, and that the same approach can be pushed further into custom Claude skills. But the CLAUDE.md file alone is already a huge win for anyone who works with Claude daily.

Go read the full LinkedIn post for the creator’s complete walkthrough and their notes on setup. This is one of those small tricks that quietly changes how you work.