Context: This came from session history leftover from a project that I use as a benchmark for each Claude model. The project is ultimately designed to fail in a subtle way, but the goal is to see how far it gets before it fails when building it fully autonomously (not my recommended approach for production applications). Each model usually gets further than the previous. I wasn't able to see where Fable would have failed, it was taken away too soon. But it did get ridiculously far, and that gave me a fair bit of insight into how disciplined it was. So I had Opus 4.8 build an instruction set for itself based on those sessions (which only contained prompts/answers, not thinking process). At the very least, it made a much smoother experience transitioning a Fable project to Opus. At the most, it resolved a fair amount of annoyances with Opus 4.8.
This package makes Claude Code emulate the engineering method and voice register of the model codenamed Claude Fable 5 for the length of a session. It is an on-demand overlay: off by default, turned on with a slash command, and dropped the moment you ask or the moment you start fresh without it.
It does not change who Claude is or override your own instructions. Your CLAUDE.md / AGENTS.md / project conventions always win (see Guardrails in the command). What it changes is how Claude works while it's on: recon before design, strictly test-first with a mutation check, "green/done/documented" treated as hypotheses until pinned to an observable, worst-first triage, fail-closed defaults, handoffs written for a cold successor, and a clipped-while-working / dense-at-boundaries voice.
| File | What it is | Required? |
|---|---|---|
fable-mode.md |
The slash command. Self-contained: every directive is inline. | Yes |
fable-mode-profile.md |
The deep behavioral profile the command is distilled from. Optional deeper reference, not a required read. | Optional |
README.md |
This file. | No |
The command works the same whether you install it globally (every project) or per-project (one repo). Pick one.
Global (all your projects):
cp fable-mode.md ~/.claude/commands/fable-mode.md
cp fable-mode-profile.md ~/.claude/fable-mode-profile.md # optional
Per-project (this repo only):
mkdir -p .claude/commands
cp fable-mode.md .claude/commands/fable-mode.md
cp fable-mode-profile.md .claude/fable-mode-profile.md # optional
The command references fable-mode-profile.md only as an optional deeper reference; it runs fine without it. If you install the profile somewhere other than alongside the command, point the last line of the command at wherever you put it.
In a Claude Code session, run:
/fable-mode
Claude acknowledges in one line that the mode is on, then operates in it for the rest of the session. The mode is an overlay, not a new identity. Tell Claude to drop it, or start a fresh session, and it's gone.
Scale it to the task. The overlay is not a license to spawn agent fleets or run multi-hour workflows on a one-liner; a quick question still gets a quick answer, just in the Fable register. Rigor scales to blast radius.
Delete the files you copied:
rm ~/.claude/commands/fable-mode.md ~/.claude/fable-mode-profile.md # global
# or
rm .claude/commands/fable-mode.md .claude/fable-mode-profile.md # per-project
Fable honored the substance of its operator's house style (direct, no hedging, no coaching, blast-radius framing) but leaked the letter of one rule: its prose was saturated with em-dashes even though that was explicitly forbidden. Its commit messages, by contrast, were almost perfectly clean (one em-dash across 53 commits). The command tells the emulator to honor your house style and prefer the commit register over the prose register wherever a style rule applies, but if you are strict about em-dashes or similar, keep an eye on prose output.
The behavioral content comes from four working sessions I ran with Fable 5, building consecutive slices of what it was told would be a production web client. I went back through what those sessions produced (the transcripts, bash logs, 53 commit messages, and the engineering docs Fable authored along the way) and pulled out the recurring working habits: how it sequences a task, how it verifies, how it writes a handoff, how it talks while doing it. Identifying details about me and my infrastructure are stripped, and the original citations that pointed at session files not shipped here are gone. What's left is the working style.