AI-iness density pre-check to avoid over-correcting human-first text

[adelaidasofia]

Problem

When humanizer runs on text that's already mostly human-written (personal journals, rough drafts, meeting notes), it can over-correct. Fragments, first-person voice, and specific names get flagged by lower-tier rules even though the text isn't AI-generated. The result is unnecessary rewrites that strip voice from already-authentic writing.

Proposed solution

A pre-flight density check that counts Tier-1 AI tells per 100 words before applying any fixes, then selects a pass strength:

• Low density (0-2 tells/100 words): Light mode. Only apply dead-giveaway patterns (Tier 1). The text is human-first; leave it alone except for obvious AI artifacts.
• Medium density (3-5 tells/100 words): Mixed mode. Apply Tiers 1-2. Some AI assistance detected but the voice is mostly authentic.
• High density (6+ tells/100 words): Full mode. Apply all tiers. This is AI-first text that needs comprehensive humanization.

Why this matters

The current behavior is all-or-nothing: every rule runs on every text. This works great for fully AI-generated content but creates false positives on human-authored or lightly-assisted text. The density check makes humanizer safe to run on any input without worrying about destroying authentic voice.

I've been running this in my fork (adelaidasofia/humanizer) and it eliminates the "humanizer made my writing worse" failure mode on personal/journal content.

Happy to send a PR.

[voidborne-d]

This is a great proposal. The over-correction problem is real — I've seen it ruin perfectly good personal writing by stripping out the exact quirks that made it sound human.

Your density threshold approach maps well to what we built in humanize-chinese for Chinese text. We score 0-100 based on feature extraction (connector density, sentence structure uniformity, idiom usage rates, etc.) and then gate the rewriting intensity on that score. In practice the tiered model works much better than all-or-nothing:

• < 30: skip rewriting entirely, text is already human-sounding
• 30-60: light touch — only fix the most obvious AI patterns
• 60+: full rewrite with style transforms

One thing we learned: the per-100-words window matters a lot. AI-generated text often has uneven density — the intro and conclusion are AI-heavy while middle sections (where the author inserted their own analysis) read fine. A global average misses this. We ended up scoring paragraph-by-paragraph and only rewriting the high-scoring blocks.

If you go with the tiered approach, consider making the thresholds configurable. Different domains have very different baselines — academic writing naturally hits more "AI tells" (formal connectors, hedging language) than blog posts, so a fixed cutoff produces false positives on scholarly text.

+1 on the PR. Would be a solid improvement.

[adelaidasofia]

Pinging on the PR offer for this one too. The density pre-check is running in production in my fork (adelaidasofia/humanizer) on journal and rough-draft content -- it's solved the over-correction problem cleanly. Ready to submit whenever you want to look at it.

[Seithx]

SKILL (1).md

here is the update skill.