Fable 5 Agents

Three production Claude Code subagents, mined from a 5-day window with Fable 5 — and preserved to run on Opus, forever.

For five days I had access to Fable 5. In that window it did real work across my stack — it hardened my infrastructure, it built browser games and verified them itself, it diagnosed why my Mac kept rebooting. Then the access was gone.

So I did something about it: I went back through every session Fable 5 ran, extracted the method — the exact steps, the tools, the gotchas it learned the hard way — and rebuilt each one as a Claude Code subagent that runs on Opus. The method outlives the model.

This repo is three of those agents — the ones that are useful to anyone, cleaned of anything specific to my accounts. They're battle-tested methods, not toy examples.

---

The agents

Agent	What it does
web-security-hardener	Audits a Cloudflare Tunnel's ingress, gates sensitive hosts behind Cloudflare Access via the API (zero downtime — never a cloudflared restart), adds a hardened HTTP security-header set, and verifies every host with curl. Keeps your public allowlist open.
arcade-game-builder	Builds/upgrades browser games (or any interactive web page) and browser-verifies each one via its own ?selftest=1 bot — a self-playing script that prints a PASS/FAIL verdict — using Claude_Preview before the typecheck → build → deploy → 200 gate. Never ships a black screen.
mac-crash-disk-doctor	Read-only-first macOS stability triage: df first → watchdog-panic / crashloop diagnosis from panic + .ips logs → guarded launchd/brew remediation. Built to answer "why does my Mac keep randomly restarting?" without guessing.

Each agent carries its full doctrine, its ordered workflow, and the specific gotchas it learned (a cloudflared restart drops every site at once; setTimeout-paced self-test bots false-hang in background tabs; a leading dot does not disable a launchd job). That hard-won detail is the point.

---

Install

Option A — as a Claude Code plugin (recommended)

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

Option B — manual (always works)

Copy the agent files into your Claude Code agents directory:

git clone https://git.dajai.io/DajaiStewart/fable5-agents.git
cp fable5-agents/agents/*.md ~/.claude/agents/

Then spawn one:

Agent({ subagent_type: "web-security-hardener", prompt: "audit my tunnel and gate anything sensitive" })

Some agents use MCP tools (claude-in-chrome for dashboard confirmation, Claude_Preview for browser verification). They degrade gracefully if a tool isn't connected — the core method runs on Bash/Read/Edit.

---

Why "Fable 5"?

Fable 5 was a frontier model I had for a very short time. These agents are how I kept what it taught me. There are 13 in the full set; the other 10 are wired to my own revenue and infrastructure and stay private. These three are the ones that travel.

If you want the method behind the method — how you mine a model's own session transcripts and turn them into permanent agents — I wrote it up here: The full story.

Hosted on my own sovereign git hub — git.dajai.io — not GitHub. Built by DAJAI Stewart · dajai.io · MIT licensed. Use them, fork them, sell what you build with them.