The AI-Voice Filter

# The AI-Voice Filter

Five diseases decide whether your writing reads as yours or as a machine's. You cannot hear them in your own draft — that is the whole problem.

Blocklists rot. Chasing "don't sound like AI" gives you a beige voice, because a negative has no target. So this filter does not chase a negative: it names five specific diseases, tells you the fix for each, and flags every hit on the exact offending line before it rewrites anything. The goal is not to sound un-robotic. It is to sound like a specific person who has thought about the thing.

These are the five, verbatim from the filter's own registry — the count in this article, in the video, and in the file are the same number because they are the same list.

1. Vagueness compression — it describes a category, not a thing. *"Users were frustrated"* becomes "users clicked export six times because nothing loaded."
2. Significance inflation — it treats a normal fact like a turning point. *"This marks a pivotal shift in onboarding"* becomes the fact, stated, with the reader left to weigh it.
3. Hedged confidence — it has a position but will not commit. *"It could be argued that…"* becomes the position, or the sentence gets cut.
4. Rhythmic flatness — every sentence the same length, every paragraph three sentences. The fix is to break the meter.
5. Borrowed authority — it sounds like a McKinsey deck or a LinkedIn thinkfluencer, with no fingerprint. The fix is to write it the way you would say it to one person.

### Five diseases, six flags — why both numbers are right
The worked example in the kit audits one paragraph and reports six flags; this article and the video count five diseases. That is not a discrepancy, and it is worth understanding because it is how the audit actually reads: a disease is a class, a flag is one offending phrase. One paragraph can trip the same disease twice, and one phrase can be filed under a finer label ("dead opening", "metaphor", "vocab tells") that rolls up into one of the five. Same audit, two grains.

This is the "before" from the kit, unedited:

> In today's fast-paced landscape, it's important to note that great onboarding isn't just about features — it's about the journey. By leveraging a seamless, robust flow, teams can unlock transformative retention and truly elevate the user experience.

Thirty-six words, and every one of the five is in there:

• Borrowed authority — *"In today's fast-paced landscape… by leveraging a seamless, robust flow."* Six vocab tells in one sentence. This is the loudest one, so it dies first.
• Vagueness compression — *"great onboarding"*, *"the user experience"*, *"teams"*. Every noun is a category. There is not one number, name, or lived detail anywhere in the paragraph.
• Significance inflation — *"transformative retention"*, *"truly elevate"*. Nothing behind either.
• Hedged confidence — *"teams can unlock."* Can. It states a capability instead of a result, which is the same evasion as *"it could be argued that"* wearing a different hat.
• Rhythmic flatness — both sentences open on the same comma'd adverbial phrase (*"In today's fast-paced landscape,"* / *"By leveraging a seamless, robust flow,"*) and land two words apart in length. Same shape twice.

And the "after", which is the same idea in a voice:

> Most users quit onboarding on step 2. Ours asked for a credit card before it showed the product, so 42% dropped there. We moved the card to the end and the drop fell to 9%.

The dead opening is gone. The *"not X, it's Y"* reframe is gone. Every buzzword is gone. And the vague claim (*"transformative retention"*) became a checkable one: 42% to 9% on step 2. Same idea. Now it sounds like a person who actually watched the funnel.

• `ai-voice-filter.md` — the filter itself, as a drop-in Claude / Claude Code skill: the five diseases, the specificity ladder, the negative-parallelism ban, the four blocklists, the audit pass.
• `AUDIT-CHECKLIST.md` — the nine-step final pass, standalone. Run it on any draft in two minutes.
• `EXAMPLE-before-after.md` — the paragraph above, filtered, with every tell named, so you can see exactly what it catches.
• `README.md` — install in two minutes.

1. Drop ai-voice-filter.md into your skills folder as anti-ai-writing/SKILL.md (~/.claude/skills/anti-ai-writing/SKILL.md), or paste it into any chat as a system rule.
2. Finish a draft normally first — composing against a blocklist produces stilted output. The filter is a final pass, not a writing method.
3. Say "run the AI-Voice Filter on this" (or "anti-AI audit this"). It returns each tell flagged with the exact offending line, then a rewrite in your voice.

No API key, no account, no setup. It re-runs on every future piece — install once, and the tells you fixed this week do not come back next week, because the filter is doing the checking instead of your memory.

## What it will not do

It forces one voice, so applied literally it will sand off a quirk you meant to keep. That is a real cost, not false modesty — and the filter says so about itself: *spirit beats letter; if a rule makes the sentence worse, break it.* Do not avoid a banned word that is the exact right word. The filter finds the machine-made parts; deciding which of your oddities are load-bearing stays yours, and it always will.

Comment VOICE on the reel and I'll send it over, or just grab it from the button above, no comment needed. Nothing expires and nothing is held back: the same five diseases, the nine-step checklist, and one real paragraph filtered end to end. Pay the two-minute install once; every draft after that runs through the same audit before anyone reads it.