English | [ํ๊ตญ์ด](README.ko.md)
# fablize โ run Opus like Fable
[](https://github.com/fivetaku/fablize/stargazers)
[](https://opensource.org/licenses/MIT)
A Claude Code plugin that makes Opus (or any Claude model) **see a task through to the end โ with evidence and verification โ 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** โ 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 โ actually running what you build, seeing it through, investigating systematically โ 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) | โ
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 | โ
shipped | A procedure โ decompose, checkpoint, refuse completion without proof. |
| Systematic investigation (reproduce โ hypotheses โ causal chain) | โ
shipped | A procedure โ on par with what the strong model already does. |
| Early-stop prevention | โ
shipped | A deterministic hook โ catches "I'll do X" without doing it. |
| Out-of-spec defect discovery | โ not possible | Capability. Injection was refuted โ the model finds it, or it doesn't. |
| Open-ended creative detail | โ not possible | Capability. Shows only where there is no fixed answer. |
| Self-driven propagation depth | โ 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** โ render/executable artifacts (HTML, SVG, games, charts) are run and observed before completion.
- **Multi-story verification gate** โ `goals.py` decomposes work and refuses a groundless "done."
- **Investigation protocol** โ reproduce, compete hypotheses, trace the full causal chain.
- **Early-stop hook** โ blocks promising-without-doing.
- **Per-task router** โ 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 โญ on the repo means a lot โ 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 โ decompose + verification gate. Debugging โ investigation protocol. Render artifact โ verification grounding. Hard task โ 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 โ 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") โ phrasing offers as questions avoids it.
## License
MIT โ see [LICENSE](LICENSE).