fablize

本页保存的是公开项目资料快照，阅读过程不需要连接 GitHub。

English | 한국어

fablize — run Opus like Fable

图片：GitHub stars 图片：License: MIT

A Claude Code plugin that makes Opus (or any Claude model) see a task through to the end — with evidence and verification — as procedure, not as luck.

Why

When Fable 5 shipped, I ran a controlled comparison of Fable 5 and Opus 4.8 (an A/B set of 19 runs plus 26 real working sessions, ~1,500 tool calls). The finding:

• On closed, answer-bearing work (code, logic, builds), the two models were effectively tied.
• The gap appeared only on open-ended work, and its nature was “following an implication one step further.”
• That depth is model capability — it could not be transferred by instructions or a harness. An injection experiment confirmed it: Opus could not reproduce the defects Fable found on its own.
• But the procedure of good work — actually running what you build, seeing it through, investigating systematically — does transfer.

fablize applies only the procedures whose effect was verified. It does not raise the model’s ceiling; it makes the model reach its own ceiling.

What transfers and what doesn’t

Trait	Transferable?	Reason
Verification grounding (run & observe the artifact)	✅ shipped	A procedure Opus skipped “not because it can’t, but because it didn’t.” Injecting it raised render-verification behavior measurably.
Multi-story completion + evidence gate	✅ shipped	A procedure — decompose, checkpoint, refuse completion without proof.
Systematic investigation (reproduce → hypotheses → causal chain)	✅ shipped	A procedure — on par with what the strong model already does.
Early-stop prevention	✅ shipped	A deterministic hook — catches “I’ll do X” without doing it.
Out-of-spec defect discovery	❌ not possible	Capability. Injection was refuted — the model finds it, or it doesn’t.
Open-ended creative detail	❌ not possible	Capability. Shows only where there is no fixed answer.
Self-driven propagation depth	❌ not possible	Capability. Directed propagation transfers; self-started depth does not.

The non-transferable rows are the model’s job (or a human’s), not a harness’s. When you hit them, fablize tells you to escalate instead of pretending.

What’s included (verified only)

• Verification grounding — render/executable artifacts (HTML, SVG, games, charts) are run and observed before completion.
• Multi-story verification gate — goals.py decomposes work and refuses a groundless “done.”
• Investigation protocol — reproduce, compete hypotheses, trace the full causal chain.
• Early-stop hook — blocks promising-without-doing.
• Per-task router — injects only the matching verified discipline.

Negligible or unverified ideas (style mimicry, broad reasoning injection, a silent-recovery guard, a review-recall scan) are not shipped. They stay in personal development until a controlled experiment confirms their effect.

Install

/plugin marketplace add fivetaku/fablize
/plugin install fablize

The per-task router (a UserPromptSubmit hook) registers automatically.

For always-on operation (the rules resident in context), run once:

bash ${CLAUDE_PLUGIN_ROOT}/setup/setup.sh    # choose local (recommended) or global

Uninstall: bash ${CLAUDE_PLUGIN_ROOT}/setup/uninstall.sh

If fablize helps, a ⭐ on the repo means a lot — setup asks once (a single opt-in question); it never stars without your explicit “yes”.

How it behaves

• Trigger with /fablize or phrases like “see it through”, or automatically when installed always-on.
• 2+ stories → decompose + verification gate. Debugging → investigation protocol. Render artifact → verification grounding. Hard task → adaptive thinking plus an /effort xhigh suggestion.
• At the capability ceiling, escalate to a stronger model or a human.

Honest limits

• It cannot raise model capability. Open-ended creative quality and self-driven discovery are out of reach — that is a model-choice decision, not a harness one.
• The effect numbers come from a small, single-family (Claude) self-measurement (the 19-run A/B set). The direction is solid; the decimals are not asserted.
• The early-stop hook can misfire on a declarative offer (“I’ll write the report if you want”) — phrasing offers as questions avoids it.

License

MIT — see LICENSE.