⚗️fablize
Makes Claude Opus work like Fable 5: it finishes the job, checks its own work, and won't say "done" without proof.
Installation
- Make sure Claude Code is installed.
npm i -g @anthropic-ai/claude-code - Copy the block below and paste it into your terminal.
The lines run one after another — everything the plugin needs, in one paste.
- Restart Claude Code and use it.
Quit and reopen Claude Code so the plugin loads. Its commands and skills are then available in every session — run
/pluginto confirm it's installed. - Don't skip — turn it on
Once Claude Code is open, type
/fablize:setupand pick "Local — this project only".That's the on switch: from then on, fablize is always working — you never have to think about it again. Skip it, and fablize stays off until you type
/fablizein a chat.Advanced — the same switch as a terminal command:
bash ${CLAUDE_PLUGIN_ROOT}/setup/setup.sh
Prefer to read the source first? Open on GitHub.
What it does
Fable 5 — Anthropic's newest model — is known for one thing above all: it finishes jobs properly. It runs what it builds, looks at the result, and never says "done" without proof. fablize teaches those habits to the model you already use.
After you install it, Claude will:
- Finish the turn — no more replies that end with "I'll do X next" without actually doing X
- Prove it works — it runs or opens what it just built and checks the output before claiming success
- Debug properly — bugs get reproduced and traced to the real cause instead of patched on a guess
One honest note: this can't make a smaller model as smart as Fable 5 — it makes it as disciplined. The author compared Fable 5 and Opus 4.8 across ~1,500 real tool calls and kept only the habits that measurably transfer. The full analysis is on GitHub.
Works with
- Claude Code — this is a Claude Code plugin; it runs inside it.
- Any Claude model — most useful on Opus and Sonnet.
- Nothing else needed — no API keys, no accounts; everything runs on your computer.
What's inside
- Early-stop guardhook
Catches replies that end with "I'll do X next" — and makes the model actually do X first.
- Verification grounding packpack
Makes the model run or open what it just built and look at the result before saying it works.
- Investigation protocol packpack
For bugs: reproduce it, list the suspects, prove the cause — then fix it.
- Per-task routerhook
Detects what kind of task you gave and switches on the right checks automatically.
- /fablize:setupcommand
Type this once after installing to switch the plugin on (it asks: this project, or all projects).