I Had Fable 5 for 5 Days. I Turned It Into 13 Agents That Never Expire.

For five days, I had Fable 5.

It was a frontier model, and for a short window I could point it at real problems. It didn't write toy examples. It hardened my infrastructure. It built browser games and then verified them itself before shipping. It diagnosed — from kernel panic logs — exactly why my Mac kept restarting in the middle of the night, and fixed it.

Then the access was gone.

Most people would shrug and move on. I don't lose things that work. So I did something that turned out to matter more than the five days themselves: I kept the method.

The insight: a model's real value is in its transcripts

When you work with a model inside a coding harness, every session is logged — every command it ran, every file it touched, every dead end it hit and how it got out. That log is not a chat history. It is a recording of a working method.

Fable 5 was gone, but the recordings weren't. Across those five days it had run eight distinct workstreams. A second, later window added more. Buried in those transcripts was everything that made the model useful: the exact ordered steps, the tools, and — most valuable of all — the gotchas it learned the hard way.

So I mined them. For each workstream I pulled out the goal, the tool chain, the commands, the failures and their fixes, and the hard rules. Then I rebuilt each one as a Claude Code subagent running on Opus — a small, focused operator that is that method, preserved exactly.

Thirteen of them. The model expired. The method didn't.

Why this beats "just prompt a fresh model"

A blank model is brilliant and amnesiac. Every session it re-derives the same lessons, re-hits the same walls, and sometimes re-breaks the same things. A method-agent starts where the last expert left off.

The difference lives in the details a fresh model doesn't know yet:

• Restarting a Cloudflare tunnel to change one rule drops every site behind it at once — so you make the change through the API instead, with zero downtime.
• A self-playing game-test bot paced with setTimeout falsely reports a hang in a background tab, because browsers throttle background timers — so you drive it with microtasks.
• A leading dot on a launchd plist does not disable the job — launchd loads any .plist, hidden or not — so a "disabled" service keeps crashlooping and eating your RAM until the machine panics.

None of that is in a fresh model's head. All of it is baked into the agents. That is the whole point.

Three of them are yours, free

Ten of the thirteen are wired to my own revenue lines and private infrastructure. Those stay private — they're my edge.

But three of them are pure method, useful to anyone, and I cleaned them of anything specific to my accounts. They're on GitHub, MIT licensed, installable as a Claude Code plugin:

• web-security-hardener — audits a Cloudflare Tunnel, gates sensitive hosts behind Cloudflare Access via the API with zero downtime, adds a hardened HTTP security-header set, and verifies every host with curl.
• arcade-game-builder — builds and upgrades browser games, and browser-verifies each one through its own self-test bot before the typecheck, build, and deploy gate. It never ships a black screen, because it looks before it says "done."
• mac-crash-disk-doctor — read-only-first macOS stability triage. Disk first, then it reads the panic logs and tells you the real reason your Mac keeps rebooting, then remediates under guard.

Install:

# in Claude Code
/plugin marketplace add https://git.dajai.io/DajaiStewart/fable5-agents
/plugin install fable5-agents

Or clone it and drop the files into your ~/.claude/agents/ folder. That's it. And no, it's not on GitHub — it's on my own git hub, git.dajai.io, which I stood up for exactly this.

The bigger idea

Frontier models will keep arriving and keep changing. Access comes and goes. What you actually want to keep is not the model — it's the way of working you discovered with it. Transcripts are that way of working, sitting in plain text on your disk, waiting to be turned into something permanent.

I turned five days into thirteen agents. You can do the same with any good session you've ever run.

The three above are the free ones. There's a fourth I'll never open-source — a mastering agent that runs my music through a specific chain on my own rig and delivers a finished master. That one doesn't get sold as a file. If you want your record run through it, that's a different conversation.

Everything else — the repo, the method, the code — is at dajai.io and on my own hub at git.dajai.io. Take it and build.

— DAJAI Stewart