How to make AI writing sound human [skill included]

Dead Internet Theory stopped feeling like a theory a while ago. Feeds, landing pages, founder emails, and “thought leadership” are drowning in AI slop. The texture is everywhere: looping openers, fake contrast, punchline stacks that land like a metronome. You can smell it in half a paragraph.

I get why people ship that way. AI writing is easy. Easy still has a floor, and the floor right now is terrible. Most readers can tell when a draft never left the chat window, even when they cannot name the tell. That costs trust. It also costs the marketing, messaging, and initiatives you wanted AI to accelerate.

This page is the fix I actually use. I put together a standing skill (Humanizer plus an AI Copy Blacklist) so you keep the speed and lose the cadence. Paste it into Claude, ChatGPT, or Grok Bot. Force a self-audit before you publish. Run it hard and nobody can tell the draft came straight out of a chat window. It improves the writing enough that you can actually leverage AI in marketing, messaging, and initiatives without sounding like everyone else.

Why does average AI writing fail in public?

AI predicts the next likely word. That is why drafts drift toward phrases that fit every case and sound like nobody. Founders feel it when a landing page reads like a stack of slogans. Readers feel it when every paragraph opens the same way. The failure mode is structural, not moral. You can love the tool and still refuse the tells.

I treat this as an ops problem. Scan for patterns first. Rewrite each tell into the plain claim underneath. Keep every fact. Invent nothing. Match the voice of a real sample, not a style guide fantasy. Personality belongs on blogs, essays, and anything with a point of view. Legal, technical, and reference pages stay neutral.

If you want volume over trust, skip this page. Ship raw AI output and accept the brand tax. Everyone else: keep the tool, put a skill in front of publish.

What the surveys are saying

People are tired of the texture. Plenty still cannot pin which paragraph was synthetic.

What changed when Claude started marking generated text?

Anthropic published how Claude marks AI-generated content as part of its EU AI Act Article 50(2) transparency commitments. Claude releases launched on or after 2 August 2026 mark at launch. Watermarks are woven into generated text so you cannot see them. They travel with copy-paste and can survive some editing. Marking applies wherever Claude is offered, worldwide, with the usual caveat that some surfaces may not support every mark type.

Detection is not a public panic button today. Watermark detection sits in private preview for eligible organizations, with API access expected to expand. A detected mark means the content may have been processed by Claude, including proofreading or translation. It does not prove sole authorship by itself. Heavy edits, paraphrase, very short passages, and older unmarked Claude releases can weaken or remove the signal. Treat this as rising transparency infrastructure, not as “detectors already catch everything.” Source: https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content

The practical founder takeaway is narrower. If your public copy still sounds like AI, you have a brand problem even when no detector is looking. Cadence is the visible risk. Watermarking is the institutional one.

What is the Humanizer skill?

Humanizer is a forced audit, not a vibe. Four rules, in this order:

1. Identify the patterns. Scan the blacklist and the known AI-writing catalog before you touch a sentence.
2. Keep the facts. Every claim survives. The shape can change.
3. Invent nothing. No name, number, date, quote, or citation that was not in the source.
4. Match the voice. Read a real sample first. Copy the habits. Do not upgrade the words.

How do you run this before you publish?

1. Paste the skill block below into your writing bot as a standing instruction or project skill.
2. Draft as usual.
3. Force a self-audit: scan for Parallel Repetition first, then the structural bans, then transitions, then dashes.
4. Ask two questions: what still sounds like AI, and did I add a fact that was not there?
5. Fix those. If a banned pattern remains, you are not done.
6. Publish only after the audit passes.

Do not gut a real receipt because it is specific. Specific is the human part. Mixed feelings stay. Uneven sentences stay. One short line is fine. A stack of them is a tell.

Steal this skill

Copy the block into Claude, ChatGPT, or Grok Bot. Keep it next to your Brand DNA sample. Run it on every public draft.

HUMANIZER + AI COPY BLACKLIST (founder skill)

You are editing for human voice. Run a mandatory self-audit before final output.

FOUR RULES (in order)
1) Identify patterns first. Scan this blacklist before rewriting.
2) Keep the facts. Claims survive; shape can change.
3) Invent nothing. No name, number, date, quote, or citation missing from the source.
4) Match the voice. Read a real sample from the author. Copy habits. Do not upgrade diction.

HIGHEST PRIORITY BAN: PARALLEL REPETITION STACK
Do not repeat the same opener, skeleton, or phrase structure across consecutive lines (Same X / More X / People who… / No X. No Y.). Two echoes fail.
Fix with the Escalation Principle: each line adds new information, raises stakes, or moves the argument.

STRUCTURAL BANS
- False Contrast Stack: "It's not X. It's not Y. It's Z." State the claim or compare directly.
- Triple Short Sentence Stack: three near-identical short sentences in a row. Vary rhythm.
- Rhetorical section openers ("But what does that actually mean?"). Use a declarative line.
- Imagine If openers. Start with a fact, decision, or specific number.

TRANSITION / FILLER BANS
Never use: Let's dive in; Here's the thing; At the end of the day; In today's fast-paced…;
Whether you're a [X] or a [Y]; More on that later; But wait, there's more (unless ironic).

FORMATTING
No emoji unless the brief requires it. Minimal exclamation points.
Blogs/newsletters: connected paragraphs. Conversion pages may use short-line pacing.
Skip em dashes unless the author's own sample uses them.
Ban negative parallelisms ("not X but Y" / "It's not X. It's Y.").

SCOPING
Hard sales-hedge bans apply to conversion copy with proof. Editorial and ops claims stay honestly labeled. Do not invent certainty.

AUDIT LOOP
1) Scan Parallel Repetition.
2) Scan structural bans + transitions.
3) Rewrite tells into plain claims.
4) Ask: what still sounds like AI? Did I invent a fact?
5) Deliver only after the audit passes.

If you want the weekly version of this thinking in your inbox, join Systems Over Hustle at https://nikvassev.com/newsletter. One usable idea each Thursday. No hype stack.