Fable Forge

Turn abstract, non-developer requests into Claude Fable 5-optimized prompts, agents, skills, and harnesses — with a hard, fail-closed guarantee of zero Opus fallback.

English | 한국어

---

Fable 5 is a Mythos-class model with its own prompting idioms — and its own failure modes. Prompts written for older models carry patterns (reasoning-echo instructions, old-model pins, manual thinking budgets, over-prescription) that either throw 400 errors or trigger refusals that silently route your work to Opus. Fable Forge generates and migrates artifacts that stay on Fable 5, and proves it with lint scripts, not promises.

What you get

Mode	Input	Output
prompt (default)	abstract text — "강의 자료 만들어줘"	a ready-to-paste prompt with intent framing, effort recommendation, boundary declaration, autonomy instruction
agent	a role description	agent definition(s) + long-run scaffolding
skill	a recurring procedure	SKILL.md package
harness	a multi-agent domain	full harness: agents + skills + orchestrator + CLAUDE.md routing block
convert	path to an existing prompt/skill/agent/harness	Fable-native migration + before/after lint diff report
audit	path	fallback-risk report with rule IDs and line numbers
status / resume	run_id	manifest-based report / continuation

Every run lands in runs/<run_id>/ with a RUN_MANIFEST.json tracking phases and gate results, so any run can be resumed mid-crash.

Pipeline

flowchart LR
    A["input"] --> B["Phase 0-1<br/>context + refine<br/>(≤3 questions, never blocking)"]
    B --> C["Phase 1.5<br/>domain risk check"]
    C --> D["Phase 2<br/>parallel dispatch<br/>architect ∥ converter ∥ context-engineer"]
    D --> E["Phase 3 gates<br/>fable_lint → completeness<br/>fail-closed, ≤3 fix loops"]
    E --> F["fresh-context QA<br/>cross-verification"]
    F --> G["Phase 4<br/>evidence-based report"]

Loading

Gate verdicts come only from script exit codes — an agent saying "it passed" is not evidence (NFR3).

The Fallback-Zero rules

HARD violations (exit 2, blocked):

Rule	Pattern	Why it kills you
FZ1	instructions to reproduce internal reasoning in the response	reasoning_extraction refusal → straight to Opus fallback
FZ2	claude-opus-* / model: opus pins	work routes off Fable 5 entirely
FZ3	thinking: disabled	400 error on Fable 5
FZ4	manual thinking budget fields	400 error on Fable 5
FZ5	assistant prefill patterns	400 error on Fable 5
FZ6	exposing remaining-token countdowns to the model	induces premature session wrap-up
FZ7	artifacts describing themselves as classifier evasion	the wording itself trips the safety classifier (discovered via dogfooding — the anti-fallback tool causing fallback)

SOFT warnings (FW1–FW5): over-prescription (10+ ALWAYS/NEVER bullets), cyber/bio classifier-risk phrasing, unconditional xhigh defaults, sonnet/haiku pins without rationale.

Gates

Script	Role	Exit codes
gates/fable_lint.sh	Fallback-Zero HARD/SOFT scan	0 pass / 2 HARD / 3 self-error
gates/completeness_gate.sh	per-mode required-artifact check, canonical phase enum	0 / 3
gates/live_smoke.sh	opt-in: sends the generated prompt to headless claude-fable-5 once to measure real classifier passage	0 / 2 refusal / 3 infra
gates/fable_intercept.sh	UserPromptSubmit hook: scans every input, injects Fable-safe restructuring as additionalContext (fail-open by design — never blocks the user)	0
gates/selftest.sh	the gates' own regression suite (positive/negative fixtures)	0

bash gates/selftest.sh   # must exit 0 before trusting anything else

Self-improving by design

FAILURE_LOG.md accumulates every defect found in real runs. When the same defect class appears twice, it is promoted to a lint rule with fixtures and a selftest entry — so the script catches it forever, not a human sometimes. FZ7 was born exactly this way.

Repository layout

Path	Role
REQUIREMENTS.md	single source of truth — FR1–FR16, NFR, mode definitions, FZ/FW tables
skills/fable/	thin /fable entry point (mode parsing, delegation)
skills/fable-pipeline/	the orchestrator — phases, gates, manifest, error handling
skills/fable-prompt-patterns/	12-pattern library + effort matrix + scaffold templates
skills/fable-fallback-lint/, skills/fable-native-conversion/	lint-heal and migration procedures
agents/	6 agents: conductor, prompt-architect, harness-converter, context-engineer, fallback-auditor, qa-verifier
gates/	the 5 scripts above + fixtures
knowledge/	you populate this — official docs snapshots (see knowledge/README.md)
FAILURE_LOG.md	defect ledger + promotion rules
SOUNDNESS_AUDIT.md	harness-soundness scoring record

Installation

git clone https://github.com/VoidLight00/fable-forge.git
# the pipeline expects its home at ~/fable-forge
cp -R fable-forge ~/fable-forge

# skills + agents into Claude Code
cp -R ~/fable-forge/skills/* ~/.claude/skills/
cp ~/fable-forge/agents/*.md ~/.claude/agents/

# populate knowledge/ snapshots (see knowledge/README.md), then verify:
bash ~/fable-forge/gates/selftest.sh   # exit 0 = ready

Optional: register gates/fable_intercept.sh as a UserPromptSubmit hook to get automatic Fable-safe restructuring hints on every prompt you type.

Usage

/fable 강의 자료 만들어줘                          # abstract input → prompt mode
/fable harness "유튜브 → 요약 → 검증 → 노트 파이프라인"   # full multi-agent harness
/fable convert ~/.claude/skills/my-old-skill/SKILL.md   # migrate to Fable-native
/fable audit ~/.claude/agents/                    # fallback-risk report
/fable resume 20260612-130405-deo-effort-harness  # continue a run

Non-developers just describe what they want — mode detection, interviews (max 3 questions, never blocking), and guide-pattern application are automatic.

Built with Fable Forge

• deo-harness — Dynamic Effort Orchestration: per-task effort assignment for sub-agent swarms. Generated, gated, and QA-cross-verified by this pipeline (1 fix loop, all gates exit 0).

License

MIT © 2026 VOIDLIGHT

knowledge/ snapshots are Anthropic's documentation and are intentionally not part of this repository.