humanize-test

Test writing for the tells that make it read machine-written. This is a diagnostic, never a rewrite. It returns a scorecard, the flagged lines, and why each one fired, then hands the fix back to a human.

The one rule: diagnose, don't rewrite

The skill flags. A human fixes.

This matters more than it sounds. Handing a flagged draft back to the model to repair produces the exact tells you were trying to remove, because the repair is generated by the same process that generated the problem. The value is in the diagnosis. If someone explicitly asks for fixes, give two or three directions per flag, never a silent full rewrite.

Inputs

1. The text.
2. The surface, which governs how strict the tests run. If it isn't given, infer it or ask.

• spoken-personal (a founder's LinkedIn or X, a personal brand, a story). Roughness is good here. All tests strict.
• clean-editorial (a company page, documentation, formal B2B). Clean is correct here. Tests 1 and 6 relax, so precision and polish are not faults. Tests 2 through 5 still apply.
• unknown. Run strict, but label roughness flags "only if this is a personal or spoken surface."

Why AI writing fails

The tests below aren't vibes. Each one is asking a version of the same question: does a real mind speak here?

• No inner monologue. A reader runs an internal conversation while reading, and strong copy drives it. Generated prose discards it, so the writing goes flat.
• It says what's expected, not what's said. A model completes the likeliest next words, so it meets every expectation and breaks none. Human writing chooses when to break expectation and when to meet it. An unbroken run of met expectations is itself the tell.
• Generalities sound generated, specifics sound lived. Exact numbers, names, dates, the one odd detail only someone present would know. If a model could have said it without being there, it doesn't count as proof.
• Show, don't claim. Write for the skeptic. Auditable detail beats adjectives, and you should assume the reader distrusts you.

The 6 tests

The cheat sheet (what test 3 hunts)

• Rule-of-three everywhere, tricolons as the default rhythm
• "It's not X, it's Y" and "isn't just X, it's Y" on repeat
• Symmetrical maxims and neat aphorisms
• Even paragraphs, every one three sentences long
• Hedges and throat-clearing: "It's worth noting," "In many ways," "Ultimately"
• Manufactured hooks: "In a world where," "Let's be honest," "Imagine if"
• Em dashes as connective tissue
• Abstract nouns doing work a concrete image should do
• "Not only but also," "That's the thing about," "Here's the kicker"
• Wrap-up bows: "At the end of the day," "One thing is clear"

A tell is a flag, not a kill. Ask whether it was earned or defaulted to. One deliberate parallel that lands is human. An unbroken habit of them is machine. The usual fix for a flagged maxim is to make it observed rather than engineered: hedge it, or turn it into a callback to something earlier in the piece.

Surface-awareness

This is the part that keeps the skill from being annoying. On a company or editorial surface, precision is the brand, so tests 1 and 6 do not fire on polish. On a personal or spoken surface, that same polish is the fault. Identical sentence, opposite verdict.

Always state which lens ran, so the reader knows what standard was applied.

Output

HUMANITY TEST - <surface> lens
🟢/🟡/🔴 1 Real-Person   - <one line, flagged text quoted verbatim>
🟢/🟡/🔴 2 Thought-Flow  - ...
🟢/🟡/🔴 3 Pattern       - <quote each tell, mark earned or kill>
🟢/🟡/🔴 4 Voice/POV     - ...
🟢/🟡/🔴 5 Impact        - ...
🟢/🟡/🔴 6 Credibility   - ...
VERDICT: <ship / one watch / fix N> + the single highest-leverage change

🟢 pass. 🟡 earned, keep it consciously. 🔴 fix. Quote the offending text verbatim so it's findable. For each 🔴 give a direction, not a rewrite.

What this does not do

• Rewrite the draft.
• Judge facts, statistics, or compliance. That's a separate fact-check pass.
• Replace a voice check. A voice pass asks "is this the right person's register." This asks "does this read machine-written." They stack, in that order.

Anti-patterns

• Flagging clean copy on an editorial surface. That's the surface lens existing for a reason.
• Treating every tricolon as fatal. Frequency is the signal, not presence.
• Returning a rewrite because it feels more helpful. It isn't, and it defeats the skill.