fablize
A Claude Code plugin that makes the model finish the job: result verification, evidence and systematic investigation enforced as procedure
Medium risk
We rate an entry medium when the tool runs code, makes network calls or reads project files. Check what exactly it does before installing.
Why this level
- Registers hooks that fire on the agent's actions
- Runs scripts and built artifacts to verify results
Install
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/fablizeDetects the agents on your machine, checks the risk and pins the version.
Other ways to install
Run one by one in the Claude Code chat
/plugin marketplace add fivetaku/fablize
/plugin install fablize@fablizeInstall the plugin: /plugin marketplace add fivetaku/fablize, then /plugin install fablize. For always-on operation run bash ${CLAUDE_PLUGIN_ROOT}/setup/setup.sh once and choose local or global.
Other ways from the author
/plugin marketplace add fivetaku/fablize
/plugin install fablizeThe per-task router registers automatically.
This is third-party code. Review the repository files before installing.
What it does
The plugin adds procedures to Claude Code that raise work discipline rather than model capability. A per-task router turns on the matching pack: verification of render artifacts, where the agent runs and inspects the output before completion; an evidence gate for multi-part tasks, where work is decomposed into steps and is not marked done without proof; and an investigation protocol for debugging through reproduction, competing hypotheses and a causal chain. A separate hook catches promises to do something without actually doing it. The author shipped only the techniques whose effect he measured in a model comparison, and states plainly that the plugin does not raise the capability ceiling and, at the limit, advises escalating to a stronger model or a human.
Who it is for. For developers who want the Claude Code agent to abandon tasks less often and verify its own output.
Good fit when
- The agent often reports completion without actually verifying the result
- You need a strict bug investigation: reproduce, hypotheses, causal chain
- A task has several parts and no step should be skipped
Not a fit when
- The task hits the model capability ceiling, where procedure does not help
- You need a quick one-off answer without extra checks and hooks
Example request
Build this HTML page and see it through: run it, verify the rendering and show evidence that it worksLimitations
The plugin does not raise model capability: open-ended creativity and self-driven reasoning depth are out of reach, which is a model choice, not a harness one. The effect numbers come from a small self-measurement within a single Claude family; the author calls the direction solid but not the exact figures. The early-stop hook can misfire on a declarative offer such as "I'll write the report if you want"; phrasing offers as questions avoids it. Some features require the always-on mode installed via the setup.sh script.
How to disable. Run bash ${CLAUDE_PLUGIN_ROOT}/setup/uninstall.sh to remove the always-on mode, then uninstall the plugin via /plugin.
Security check
- Registers hooks that fire on the agent's actions
- Runs scripts and built artifacts to verify results
README in short
The README explains the plugin's origin: the author compared the Fable 5 model and Opus 4.8 and kept only the work techniques whose effect was confirmed and transferable through a harness. The plugin includes render-artifact verification, an evidence gate for multi-part tasks via goals.py, an investigation protocol and a deterministic early-stop hook, plus a router that turns on the matching pack per task. Ideas with unconfirmed effect, such as style mimicry or broad reasoning injection, were deliberately left out. Installation goes through the plugin marketplace, with setup.sh for always-on mode and uninstall.sh for removal. MIT license.
FAQ
Does the plugin make the model smarter?
No. It adds procedures for finishing work and verifying results, but does not change the model's capability ceiling. At the limit it advises escalating to a stronger model or a human.
How is the right set of checks turned on?
A per-task router picks the pack itself: render verification for artifacts, an evidence gate for multi-part tasks, an investigation protocol for debugging.
Related
Open-source personal AI assistant on your own machine: answers in Telegram, Slack, Discord and WhatsApp, extended with skills and plugins
A self-improving agent from Nous Research with a TUI, messaging gateway, cron jobs and skills it writes itself
An open source coding agent for the terminal and desktop with build and plan modes
Open prompt library with a Claude Code plugin, MCP server and CLI: search, fetch and improve prompts and skills from an agent