scandnavik/writing-harness

This page was translated automatically. Technical details may differ from the original. View the original

22 stars Quality 54 Trend 14MIT

Over four months, we gradually adjusted the mechanism to automatically prevent AI slops, including settings and scripts between CLAUDE.md and hooks, aiming to help reduce AI slops during system operation.

Overview

A quality gate that stops an LLM from shipping "AI slop" in Chinese long-form writing. Three stations: an input gate, a mechanical regex gate, and a human-judgment gate with mandatory self-attestation. A writing quality gate to prevent AI slops (AI slops refer to the vague, piled-up quality of AI writing that clearly doesn't look human). It shifts the issue of "whether it will be written in an AI tone" to a post-instinct retouch, moving it forward to two fixed checkpoints: before writing and after drafting. Methodology, scripts, and agent hooks are given together, and others can clone them and apply them. Methodology and scripts are unrelated to agents (any agent/editor/CI can run); Automatic reminder wiring: Claude Code, OpenAI Codex CLI, NousResearch Hermes all have ready-made versions, see integrations/. Adjust for **context, but the methodology itself is unrelated to language; switching to a glossary can transfer to another context. LLMs writing long Chinese texts have a consistently bad taste: half-width punctuation trapped in Chinese, overflowing dashes, mainland Chinese terminology, binary contrast templates like "not X, it's Y," zero-carrying filler like sentence prefixes like "actually/honestly," and making up fake examples in a serious manner. All of this combined is an AI slop. Relying on prompts to write a few rules can't stop it, because the model selectively follows it, and there's no evidence afterward to prove it actually did it. This system adopts a 'gate' design: the mechanical parts are handed over to the program (exit code doesn't lie), the judgment part is forced to retain evidence (self-answer by posting answers for each question), and the rule base only increases, never decreases. When installing the script, copy the methodology, script, hook, and templates to ~/.claude/skills/writing-harness/, copy tighten skills to ~/.claude/skills/tighten/, and run a smoke test (the most basic "can be moved after installation"). It's **, hook wiring is something you have to do yourself (see below). exit 0 means pass; exit 10 lists every hit and line number. This script has zero dependencies and is pure stdlib; any editor, agent, or CI can accept it. Open methodology/writing-harness.

Overview (machine translated)

A quality gate that stops an LLM from shipping “AI slop” in Chinese long-form writing. Three stations: an input gate, a mechanical regex gate, and a human-judgment gate with mandatory self-attestation.

A writing quality gate to prevent AI slops (AI slops refer to the vague, overly cluttered, obviously not written by human style). It shifts the issue of “whether it will be written in an AI tone” to a fixed checkpoint before writing and after drafting based on intuition. It provides methodology, scripts, and agent hooks together, and others can clone them and apply them. Methodology and scripts are independent of agents (any agent, editor, or CI can run); Automatic reminder wiring, Claude Code, OpenAI Codex CLI, and NousResearch Hermes are all ready-made versions, see integrations/.

Regarding Taiwan Traditional Chinese context adjustment, but the methodology itself is unrelated to language; switching to a glossary can transfer it to another context.

What is this system addressing?

LLMs writing long Chinese texts have a consistently bad taste: half-width punctuation trapped in Chinese, overflowing dashes, mainland Chinese terminology, binary contrast templates like “not X, it’s Y,” zero-carrying filler like sentence prefixes like “actually/honestly,” and making up fake examples in a serious manner. All of this combined is an AI slop.

Relying on prompts to write a few rules can’t stop it, because the model selectively follows it, and there’s no evidence afterward to prove it actually did it. This system adopts a ‘gate’ design: the mechanical parts are handed over to the program (exit code doesn’t lie), the judgment part is forced to retain evidence (self-answer by posting answers for each question), and the rule base only increases, never decreases.

System Layer 3

Layer	File	Role
Methodology	`methodology/writing-harness.md`	S0 Input gate / S1 Mechanical gate / S2 Validation gate, three main stations
	`methodology/taiwan-writing-glossary.md`	Rule manual: Punctuation, Mainland China→ Taiwanese vocabulary, sentence patterns, and terminology in plain language
Mechanical Gate Script	`scripts/taiwan-style-check.py`	S1 main gate, scan for Class 11 hard rules (exit 0/10)
	`scripts/verbosity-check.py`	Sweep 10 bloat/redundant pattern
	`scripts/rewrite-diff.py`	Compare the draft with the real-life revision to find the rules that should be written into the glossary

This page was translated automatically. Technical details may differ from the original.

Recommended Tools

Try a different keyword or remove a filter.

Install

npx skillfish add scandnavik/writing-harness

New to MCP? Start with the mcpskill.org guide →