Fable is gone from subscriptions. The discipline isn't.
Install • Before / After • What it fixes • How it works • Usage • FAQ
What made Fable feel different wasn't only raw intelligence — it was discipline: lead with the answer, cover the whole request, act instead of asking permission, keep scope minimal, never claim something works without running it.
fable-mode distills that discipline into an always-on ruleset for Claude Code, injected into every session and every subagent. Opus, Sonnet, and Haiku pick up Fable-class habits for ~117 tokens of context.
Run these one at a time — each is its own command, not a single paste.
/plugin marketplace add tarunlnmiit/fable-mode
/plugin install fable-mode@fable-mode
Start a new session. You'll see FABLE MODE ACTIVE in the session context. Done.
"Fix the login bug"
| Without fable-mode | With fable-mode | |
|---|---|---|
| Opening | "I've completed my analysis of the authentication module. Let me walk you through what I looked at. First, I examined the middleware..." | "Login fails because the token expiry check uses < instead of <=. Fixed in auth/middleware.ts:42, tests pass." |
| Coverage | Patches the one line the ticket names | Finds root cause, checks the same bug elsewhere, verifies the fix runs |
| Execution | "I found the bug. Would you like me to fix it?" | Fixes it, then reports |
| Scope | Drive-by refactor + bonus README section | Exactly the fix. One line at the end if something else is worth doing |
| Honesty | "This should work now ✅" | "Fixed and tested" — or "written, not run" |
Five failure modes every Claude model reliably falls into — and the rule section that kills each one:
| # | Failure mode | Rule |
|---|---|---|
| 1 | Verbose preamble — narration, filler, "Great question!" | §1 Lead with the outcome |
| 2 | Missed scope — answers the literal sentence, misses what the request covers | §2 Enumerate what "done" includes, trace blast radius |
| 3 | Permission-asking — "Shall I proceed?" on reversible work | §3 Act, don't ask |
| 4 | Over-engineering — strategy patterns for one date format | §4 Simplest solution that works |
| 5 | False confidence — "should work" reported as "works" | §5 Never claim untested work is done |
Plus §6 — stronger-model consultation: on Haiku or Sonnet, at high-leverage moments (architectural decisions, stuck after two attempts, plan review) the model offers to consult a stronger model — Fable where available, Opus otherwise — incorporates the advice, and executes the work itself.
And §7 — per-model notes: which sections each model needs most (Opus/Sonnet: §1/§3/§4; Haiku: §2/§5).
Read the full ruleset: rules/FABLE.md
SessionStart hook ──► activate.js ──► injects rules/FABLE.md into session context
SubagentStart hook ──► same rules, wrapped for subagents
- No API calls, no network, no telemetry. One Node script printing a Markdown file.
- ~117 tokens always-on context cost per session.
- No state files. Toggling is conversational.
- Works on any model your plan runs — the rules adapt per model (§7).
.claude-plugin/ marketplace + plugin manifests
hooks/ SessionStart + SubagentStart → activate.js
rules/FABLE.md the ruleset (the actual product)
skills/fable/ the /fable skill
Nothing to configure — active every session after install.
| Say | Effect |
|---|---|
/fable |
What fable mode is, whether it's active |
fable status |
Active/inactive + current model |
stop fable mode / fable off |
Disable for the session |
fable on |
Re-enable |
Does this make Sonnet as smart as Fable? No. It makes Sonnet behave like Fable — the communication, scope, and verification discipline. Raw reasoning stays the model's own. §6 bridges the gap: for the hardest calls, the model offers to consult a stronger model and executes with that advice.
Will it conflict with my CLAUDE.md? Your project and user instructions always take precedence. fable-mode is baseline behavior; your rules override it.
What does it cost? ~117 tokens of session context. No runtime cost, no dependencies beyond Node (which Claude Code already requires).
Uninstall?
/plugin uninstall fable-mode@fable-mode
MIT © Tarun Gupta