Wiring the avoid-ai-writing Skill into the Article Pipeline

It started with a skill I stumbled on

A few days ago I was browsing the skill directory on Hermes Atlas and landed on conorbronsdon/avoid-ai-writing. Not something I was looking for, but the description caught my eye: the skill audits and rewrites text to strip the patterns that make prose read machine-generated. Some of the articles on this blog are agent-assisted, so the question was simple — could I use this to make the writing more pleasant for humans to read without hand-checking every sentence?

I installed it, read the SKILL.md, and it turned out to be more than a vague "sound more human" prompt. It ships a 112-entry word-replacement table plus 69 pattern categories, with three modes: rewrite (flag and fix), detect (flag only), and edit (edit a file in place). There are also voice profiles: casual, professional, technical, warm, blunt.

Why an AI detector can't be the judge

The skill honestly states one important thing: the patterns it flags are signals, not proof. A Stanford study by Liang et al. (published in the journal Patterns, Cell Press, July 2023) shows how unreliable AI detectors are. They tested 7 popular detectors on real human writing: 91 TOEFL essays (non-native English writers) and 88 essays from US 8th-graders. The result — the detectors nearly perfectly classified the US students' essays as human, but mislabeled more than half the TOEFL essays as "AI-generated," with an average false-positive rate of 61.3%.

The cause is in how detectors work: they measure perplexity, how predictable the word sequence is. Limited vocabulary = low-perplexity text = red flag. That double-penalizes ESL (English as a second language) writers — not because they used AI, but because their word choice is narrower. So when someone says "the detector says your writing is AI," that is not necessarily a fact.

The point is not "don't improve your writing," but "don't let the tool be the judge." avoid-ai-writing wins here because it acts as a quality editor, not a verdict. It doesn't give a score like "this is 87% AI." It just says "this sentence uses 'leverage' when 'use' is enough, want to swap it?"

How I wired it in without breaking the facts

On this blog, articles flow through tools/articles.py verify — it checks structure, SEO, and style before publish. I ported a subset of this skill into verify as a new gate: check_ai_writing(). But I made one important decision — detect mode, not rewrite.

The reason is practical: the technical articles here carry facts, numbers, variable names, and code. If I let auto-rewrite run in the pipeline, it risks "fixing" a sentence until it changes the technical meaning. That violates my own evolver principle: help mechanically, don't alter the angle or the facts.

So the gate only rejects fingerprints with near-zero false positives: chatbot artifacts (Tentu! Semoga membantu!), leaked chat-UI citation markup (citeturn0search1), AI-tool tracking URL params (utm_source=chatgpt.com), and empty placeholders ([Insert ...]). All of those are FAIL — the article cannot publish. English-only heuristics like "studies show" or "could potentially" are just WARN — they don't block, only inform.

The trial ran across 40 published articles (20 id + 20 en) with the --no-rollback flag, and came back clean: zero false positives. That means the gate doesn't reject healthy articles just for their writing style.

One thing I learned: the upstream word table is English-centric. For Indonesian articles, most entries (delve, tapestry, robust) never trigger because we don't write in English. What transfers cross-language is only the structural rules — too many em-dashes, excessive bold, bloated bullet lists, chatbot artifacts. So I locked the English heuristics to the en locale only, so they don't wrongly flag id articles.

If you run a blog and use an agent to write, my message is simple: don't fear the detector, but care about readability. A tool like avoid-ai-writing is useful when positioned as an editor, not as an automatic gatekeeper that rejects by default. In my pipeline it sits at the verification layer — it helps before publishing, it doesn't replace my writing.