Anbeeld/WRITING.md

What it solves

AI-generated text often follows predictable patterns that make it easy to recognize as machine-written. This ruleset targets those tells — repeated cadence, generic phrasing, parallel lists, overly neat conclusions, and meaning drift during edits — without breaking grammar or faking typos. It makes the output read more like a human wrote it while staying honest and useful.

How it works

The ruleset combines writing standards, web readability research, documentation guidance, detector-limit research, and direct observation of how LLMs behave by default. It starts by considering the medium, audience, reader need, and the text's job, then calibrates voice to fit the genre. It guards against fake specificity by requiring checkable details, preserves meaning when sentences are smoothed, and breaks template patterns by preferring ordinary words and cutting staged parts. It comes in three sizes: a full version (5400 words), a compact version (1500 words), and a mini version (~330 words). These can be used as direct instructions, as a skill in agent harnesses like Claude Code or Cursor, or as a persistent section in AGENTS.md.

Who it's for

Writers and professionals who use AI to iterate on their own drafts or turn loose thinking into publishable prose — articles, blogs, documentation, emails, marketing copy, and similar reader-facing text. It works with major agent tools and can be added as a lightweight instruction set. It is not meant for code comments, commit messages, or private notes, and it will not fool an experienced reader or advanced detector.

Highlights

• 5400-word full ruleset, 1500-word compact, and 330-word mini versions for different contexts
• Skill package for agent harnesses that loads on demand and costs minimal context
• Requires concrete anchors and blocks fake specifics like unsupported quotes or numbers
• Preserves scope, uncertainty, attribution, and conditions when editing
• Does not fake humanity — no invented typos, forced slang, or programmatic sentence-length variation