fable5-mode

fable5-mode

English | 简体中文

Make Opus 4.8 — or any Claude model — work like Claude Fable 5. A Claude Code skill plus guard hooks that give a non-frontier model Fable-5-grade work discipline: plan-gate, self-verification, and sub-agent routing, enforced mechanically.

The skill is invoked as fable-mode; its premise:

output quality = model capability × work discipline

---

Quickstart

# 1. Install the skill (honors CLAUDE_CONFIG_DIR; falls back to ~/.claude)
git clone https://github.com/cozytab/fable5-mode \
  "${CLAUDE_CONFIG_DIR:-$HOME/.claude}/skills/fable-mode"

# 2. (optional) register the enforcement hooks — merges into your settings.json,
#    resolves its own path, idempotent:
bash "${CLAUDE_CONFIG_DIR:-$HOME/.claude}/skills/fable-mode/install.sh"

# 3. In Claude Code, just ask for it — in any language:
   "use fable mode"   ·   "用 fable 模式"   ·   "do this rigorously, one-shot"

# 4. For a project you're taking seriously, turn on enforcement:
mkdir .fable
printf -- '- [ ] 1. first card (with a machine-checkable acceptance test)\n' > .fable/LEDGER.md

The skill works on its own; step 2 is optional (it turns the discipline into hard blocks). Claude replies in your language even though the skill is authored in English.

Design intent & what it changes

Why it exists. Weaker models mostly don’t fail by being dumb in the moment; they fail by process: thinking while typing and changing their mind halfway, declaring “looks right” without running anything, fanning work out with no verification, and quietly stopping mid-task. Those are process failures, and process is fixable with structure. Since the “discipline half” of frontier output is model-independent, you can hand a cheaper model the working habits of a stronger one and recover a real chunk of the gap — paying in orchestration steps instead of a bigger model.

What you actually get.

• Thinking happens in the cheapest phase (a written plan gate), not mid-implementation.
• “Done” means an acceptance command passed — not “looks right.”
• Critical output survives an adversarial refute pass before it ships.
• Context stays clean across long runs, via external SPEC/PROGRESS memory instead of a bloated transcript.
• The two rules models shirk most — write a plan before fanning out and don’t stop with unfinished work — are enforced by hooks, not hope.
• It’s honest about its ceiling: on a real capability wall (a long from-scratch derivation, holding a huge codebase at once, fine aesthetic judgment) it tells you to switch to a stronger model instead of faking it.

What it does not do. It won’t turn Opus into Fable 5. It closes the discipline gap, not the capability gap. The cost is real — more orchestration steps and slower wall-clock on small tasks — which is exactly why you don’t use it on small tasks. (Net token cost depends on the task; on rework-prone work the discipline can even come out lower.)

The six levers

#	Lever	What it forces
1	Plan Gate	Write docs/SPEC.md (requirements + approach + task cards, each with a machine-checkable acceptance test) before writing code. Concentrate thinking in the cheapest phase.
2	Small-card execution	Each card runs in a fresh context; don’t advance until its acceptance command passes; on 2 failures, escalate instead of flailing.
3	Adversarial self-check	Don’t “generate and ship.” Dispatch independent viewpoints to refute critical output; for wide-open problems, generate N approaches + judge + synthesize.
4	Real-product verification	All-green static checks ≠ it works. Run the real product end-to-end each milestone and leave evidence (screenshots, logs, test output).
5	Context hygiene	SPEC + PROGRESS are external memory. Restore state by re-reading them, not by dragging failed-attempt reasoning through a bloated context.
6	Checkpoint autonomy	Long background tasks get a watchdog and resumable checkpoints, so a hang or crash costs at most one card.

The full protocol lives in SKILL.md — the text Claude actually reads.

The enforcement layer (the part that’s more than a prompt)

Six levers written as prose still rely on the model’s honor. Four hooks turn the most-shirked rules into hard blocks:

Hook	Event	Effect
Profile Injector	SessionStart	Auto-injects the discipline, sized to the ledger state — full during an active round, a one-liner when idle or paused — plus the model-appropriate tier and open-item recovery.
Spawn Guard	PreToolUse (Agent/Task/Workflow)	Blocks a detailed spawn before a ledger exists (forces the plan gate), and blocks any spawn requesting a model stronger than the session’s — the model ceiling is mechanical, not just prose.
Fail-Streak Reminder	PostToolUse (Bash)	Advisory, never blocks: at every 3rd consecutive failing command it injects the attribution ladder (suspect the harness → prove the new code is running → only then debug the product, and fix the class via an invariant) — cures grinding on the wrong layer.
Close Guard	Stop	Blocks ending the turn while the ledger still has unchecked items — cures early stopping / spinning. Also enforces evidence-on-close: a - [x] card without an -- evidence: note blocks the stop (“report evidence, not adjectives” as a hard rule).

Plus hooks/fable_lint.py — not a hook but a one-shot CLI: checks the SPEC carries [measured]/[inferred]/[not-shown] source tags, every open card names its acceptance, every closed card carries evidence. Run it at wrap-up (or in CI): python3 hooks/fable_lint.py <project_dir>.

Design properties that make this safe to register globally:

• Opt-in per project via a .fable/ directory (searched upward, bounded at the git root). No .fable/ → the hooks pass through silently. They never touch projects you didn’t opt in.
• Fail-open — any hook error passes through (exit 0). A bug in a guard can never brick your session.
• Loop-safe — the close guard honors stop_hook_active, so you’re never trapped.
• Exemptions — small spawns (< 1500 chars) and forks are never blocked.

See hooks/README.md for the mechanics.

Install

Prerequisites: Claude Code, and python3 (standard library only — no third-party deps; needed only if you use the hooks). Hooks assume a POSIX environment (macOS / Linux / WSL) where python3 is on PATH; on native Windows the skill itself still works, but the mechanical enforcement layer is untested there — treat it as protocol-only.

Your Claude config directory is $CLAUDE_CONFIG_DIR if that variable is set, otherwise ~/.claude. Everything below derives from it, so it works no matter where your config lives.

Option A — automated (recommended; this is what an AI can run for you)

git clone https://github.com/cozytab/fable5-mode \
  "${CLAUDE_CONFIG_DIR:-$HOME/.claude}/skills/fable-mode"
bash "${CLAUDE_CONFIG_DIR:-$HOME/.claude}/skills/fable-mode/install.sh"

install.sh resolves its own location (so the hook paths are correct no matter where you cloned it), honors CLAUDE_CONFIG_DIR, and merges the four hooks into your settings.json without disturbing existing config. It’s idempotent — re-run it after moving the skill and it re-points the paths. Remove the hooks any time with bash install.sh --uninstall.

Or just tell Claude Code:

Install the skill at https://github.com/cozytab/fable5-mode by following its README.

Updating: after git pull, re-run install.sh. A new version can add hook events (e.g. a PostToolUse hook), and only the installer updates your settings.json — a plain git pull ships the files but won’t register new hooks. It’s idempotent, so re-running is always safe.

Note: hook registration (install or update) takes effect from your next Claude Code session — settings are read at session start, so restart or open a new session after running the installer.

Option B — manual

Clone as above, then merge these four entries into the hooks object of <config-dir>/settings.json (don’t overwrite the file; if you already have a hooks key, add these under it). Use the absolute path to where you actually cloned it — the ${CLAUDE_CONFIG_DIR:-$HOME/.claude} below is expanded by the shell at hook-run time:

"hooks": {
  "SessionStart": [{"hooks": [{"type": "command",
    "command": "python3 ${CLAUDE_CONFIG_DIR:-$HOME/.claude}/skills/fable-mode/hooks/fable_profile_inject.py"}]}],
  "PreToolUse": [{"matcher": "Agent|Task|Workflow",
    "hooks": [{"type": "command",
      "command": "python3 ${CLAUDE_CONFIG_DIR:-$HOME/.claude}/skills/fable-mode/hooks/fable_spawn_guard.py"}]}],
  "PostToolUse": [{"matcher": "Bash",
    "hooks": [{"type": "command",
      "command": "python3 ${CLAUDE_CONFIG_DIR:-$HOME/.claude}/skills/fable-mode/hooks/fable_fail_streak.py"}]}],
  "Stop": [{"hooks": [{"type": "command",
    "command": "python3 ${CLAUDE_CONFIG_DIR:-$HOME/.claude}/skills/fable-mode/hooks/fable_close_guard.py"}]}]
}

The skill alone is enough to trigger fable-mode by asking for it; skip the hooks if you only want the protocol.

For AI assistants installing this

When a user asks you to install fable-mode from this repo:

1. Config dir = $CLAUDE_CONFIG_DIR if set, else ~/.claude.
2. git clone the repo into <config-dir>/skills/fable-mode.
3. Run bash <config-dir>/skills/fable-mode/install.sh — it merges the hooks idempotently and resolves paths itself. If you instead register hooks by editing settings.json yourself, merge into the existing hooks object (never overwrite the file) and use the absolute clone path — do not hardcode ~/.claude if the user relocated their config.
4. Tell the user the hooks are optional and stay inert until a project has a .fable/ directory.

Use

Trigger the skill by asking for it by name — “use fable mode”, “work like Fable 5”, “rigorous mode”, or the Chinese equivalents (“用 fable 模式”, “严谨模式”). It deliberately does not auto-trigger just because a task is big or important — no surprise process tax; at most it may offer to enable itself. The other explicit path is a project armed with .fable/, where the hooks carry the discipline in automatically.

Enable the mechanical enforcement on a project you’re taking seriously:

mkdir .fable
cat > .fable/LEDGER.md <<'EOF'
- [ ] 1. first card (with a machine-checkable acceptance test)
- [ ] 2. second card
EOF

From then on, in that project: sessions auto-load the discipline and the right tier; you can’t dispatch a detailed agent without a ledger; you can’t end a turn with unchecked cards. Mark cards - [x] (done + verified) or - [~] ... -- deferred: reason to close them. To turn enforcement off, check everything or rm -rf .fable.

Project hygiene: add .fable/ to your project’s .gitignore — it’s this round’s working state, not history. docs/SPEC.md and docs/PROGRESS.md are durable project documents: commit those (unless you prefer them private).

Big project, small tasks? Enforcement is per-round, not per-keystroke: with all cards closed (idle) the guards stay quiet and quick fixes flow freely with near-zero injection. Mid-round, drop a PAUSED: reason line into .fable/LEDGER.md to do unrelated work without being nagged (the model ceiling stays active); remove it to resume the round.

Concurrency tiers

fable-mode’s concurrency isn’t a fixed number.

The multitasking rule applies in both tiers: independent tool calls are batched into one message, and independent side-tasks (searches, verification runs, bulk mechanical work) run as background subagents while the main thread keeps working — no idle waiting. The tiers only change the cap:

• Conservative (default) — cap of ≤5 concurrent subagents, a local rate-limit guardrail. Quality-critical implementation stays inline; the cap still gets filled with independent side-tasks.
• Throughput (opt-in) — dispatch parallel subagents readily, communicate async, don’t block. No protocol cap; field deployments range 10–500+. It trades more tokens for throughput and risks rate limits — so it’s enabled only when you ask, or auto-selected when the running model is Fable 5.

One-word controls — say a phrase, Claude writes a per-round directive line into .fable/LEDGER.md (auditable, never silent): “quality mode” → ROUTING: quality · “frugal” → ROUTING: frugal · “full speed / max parallel” → TIER: throughput · “back to normal” → remove it. Env FABLE_ROUTING / FABLE_MODE_PROFILE override; by default the tier follows the session model.

Model routing — three profiles, two iron rules. Solving the problem always outranks saving tokens; the profiles only tune how much safe downgrading you accept, and you pick one in plain words:

Profile	You say	Implementation cards	Mechanical work
quality	“quality mode / no downgrades / 质量优先”	session model, always	session model
balanced (default)	—	inherit; drop one tier only when tightly specified	cheap tier
frugal	“frugal / save quota / 节省模式”	default one tier down (acceptance still required)	cheapest tier

The selection is written into the ledger as a ROUTING: <profile> line (per-round, auditable; env FABLE_ROUTING also works) — never switched silently. In every profile: task decomposition, orchestration, design, debugging and all verification stay on the session model — a cheap model there poisons everything downstream; only cards with machine-checkable acceptance may run below the session model; two failed acceptances escalate the tier — capped at the session model (the top of the ladder is pulling the card back inline, never a stronger model: fable-mode exists to get Fable-5-grade results without Fable 5); and the verifier is always at least as strong as the implementer. This mirrors Anthropic’s own practice (Opus-class lead + Sonnet-class subagents, +90.2% over single-agent; inherit when unsure).

The Fable 5 habit set

Beyond the six levers, the skill transplants the concrete behaviors Anthropic documents for Fable 5 — so any model in a fable-mode project inherits them: ground every progress claim in a tool result; never end a turn on a promise you could act on; lead with the outcome; pause only where the user is genuinely needed; assessment before action; fresh-context verifiers over self-critique; model & effort routed by task (verification never downgraded); pass the why along when delegating; keep a lessons file; act once you have enough information (no re-litigating settled decisions); do the simplest thing that works (no unrequested refactors or defensive code); write the final summary as a re-grounding for a reader who saw none of the work; triage multi-part requests so no sub-ask is silently dropped; write code that blends into the surrounding file. The highest-value habits are auto-injected into every fable-mode session by the Profile Injector.

Starter skeletons live in templates/ — SPEC, LEDGER, PROGRESS, and an engine-neutral fresh-eyes verifier prompt.

How close to Fable 5 does it actually get?

Honest answer: frontier output = capability × discipline, and this skill can only transplant the discipline half — but it transplants essentially all of it, and mechanically rather than as prose promises: the plan gate, per-card acceptance, adversarial verification, evidence-on-close, the attribution ladder, and the model ceiling are enforced by hooks and lint, not hope. What it cannot give a weaker model: single very long reasoning chains, holding a huge codebase in mind at once, frontier-grade visual precision and taste. For those it degrades gracefully — decompose, best-of-N + judge, tools as ground truth — and tells you what remains uncertain instead of faking it. In practice that means: on well-specified, verifiable development work the gap narrows dramatically; on raw-capability walls it narrows the honest way — by saying so.

Why a skill + hooks, not a plugin or agent? A plugin restructure would add distribution convenience but no new enforcement capability, and we don’t ship forms we can’t verify end-to-end; an “agent” can only advise, not block. The current form installs with one clone + one script and is verified on a real machine. Plugin packaging is deferred, not rejected.

No stronger model? It degrades, never stalls

fable-mode is honest about capability walls — but “switch to Fable 5” is a dead end if you can’t run Fable 5. So a non-Fable session is automatically told not to defer hard steps to a stronger model or stall waiting for one. Instead it compensates on the model you have: decompose the wall into smaller verifiable steps, best-of-N + a judge, make tools/tests the ground truth, and flag residual risk instead of blocking. If you do have a stronger tier to hand off to, set FABLE_ESCALATION=on.

Layout

fable-mode/
├── SKILL.md              # the protocol Claude reads (the six levers, tiers, red lines)
├── README.md             # this file
├── README.zh-CN.md       # 简体中文
├── install.sh            # merge/remove the hooks in settings.json (path-resolving, idempotent)
├── templates/            # SPEC / LEDGER / PROGRESS skeletons + fresh-eyes verifier prompt
├── hooks/
│   ├── README.md         # hook mechanics, ledger format, install
│   ├── _fable_common.py  # shared helpers (stdin, upward .fable/ search, ledger parse)
│   ├── fable_profile_inject.py   # SessionStart: per-model tier + context recovery
│   ├── fable_spawn_guard.py      # PreToolUse: design gate (open card required) + model ceiling
│   ├── fable_fail_streak.py      # PostToolUse(Bash): attribution-ladder reminder on fail streaks
│   ├── fable_lint.py             # not a hook: one-shot discipline lint CLI
│   └── fable_close_guard.py      # Stop: open cards / hollow evidence → block turn end
└── tests/
    ├── test_guards.py    # spawn/close guards, model ceiling, PAUSED, evidence, fail-streak, lint
    ├── test_inject.py    # per-state injection, tiers, routing, escalation policy
    └── test_install.py   # install.sh: fresh/merge/idempotent/re-point/uninstall

Tests

No third-party dependencies:

python3 tests/test_guards.py    # opt-in detection, ledger presence, exemptions, git-root boundary, loop-safety, fail-open
python3 tests/test_inject.py    # per-model tier, env override, context recovery, JSON envelope, fail-open
python3 tests/test_install.py   # install.sh: fresh, merge, idempotent, re-point, uninstall, bad-JSON refusal

License

MIT © 2026 cozytab.

You may use, copy, modify, merge, publish, distribute, sublicense, and sell it — including commercially, and including in closed-source work. The only requirement: keep the copyright notice and the MIT permission text (i.e. the LICENSE file) in all copies or substantial portions. It’s provided “as is”, with no warranty and no liability on the author.