Strip the AI-writing tells out of any draft

01 Load the brand-voice canon before you touch a word.

The first move is to refuse to start without the voice source. A humanizer that rewrites against no canon does not remove AI voice. It invents a new one. The rewrite layer needs a specific human's word habits, rhythm, and point of view to inherit, or it produces a different flavor of generic. AI defensibility comes from non-AI moats

Voice is not decoration you add at the end. It is the thing AI cannot clone, because it comes from one person's lived experience and a specific way of seeing. Pull the canon: prior pieces by the named voice, the kill-list, the cadence rules, the words this brand never uses. If that substrate is missing or stale, stop. Refuse to rewrite. You would be guessing.

I want the ugly first draft.

· Madeleine Work, LinkedIn, 2026-04-10

AI-generated marketing materials strip out the humanity that makes content compelling.

Madeleine's point is the operating principle for this whole play. AI takes the essence of a message, polishes it, and strips out the humanity. The raw take is the asset. The humanizer's job is to keep the rough human edge while removing the machine residue, not to sand both off. Even OpenAI knows AI can't replace human strategic thinking for content strategy.

02 Run rule-pack detection. Flag every violation with line, match, and fix hint.

Detection is layer one. Vale-style rule packs scan the draft and flag each violation with its line number, the matched string, a severity, and a fix hint. This is mechanical work a tool does well, and it is where you start, the same way an eval shop starts with data analysis on real output, not with a theory of what could be wrong. Evals are data analysis, single judge, binary rubrics, error analysis first

"It really doesn't have to be scary or unapproachable. It really is, at its core, data analytics on your LLM application."

· Hamel Husain & Shreya Shankar, Lenny's Podcast, 2026-04-28 · “Evals are systematic data…”

The eleven rule packs, from most to least severe:

Negative parallelism is the big one because it is the single most reliable AI tell, and it survives a casual read. Detection alone catches the gap. It does not close it.

03 Score each failure mode binary. The output either has the tell or it does not.

Do not grade voice on a 1-to-5 scale. A Likert score hides the decision and produces an average that looks reasonable while the worst cases slip through. Each rule pack is a binary judge: the kill-list word is present or absent, the sentence is negative-parallel or it is not. Build narrow checks per failure mode, validate them against human-labeled drafts, and look at the cases where the check and the human disagree. Verification, not execution, is the irreplaceable human job

"1-2-3-4-5 is a weasel way of not making a decision."

· Hamel Husain & Shreya Shankar, Lenny's Podcast, 2026-04-28 · “Build LLM-as-judge as binary…”

The same card carries the consequence of getting this wrong:

"When people lose trust in your evals, they lose trust in you."

· Hamel Husain & Shreya Shankar, Lenny's Podcast, 2026-04-28 · “Build LLM-as-judge as binary…”

A humanizer that flags false positives gets ignored inside a week. Keep the judges few, narrow, and binary, and retire a pack once the model stops producing that tell.

04 Rewrite the failing spans with the voice canon and the surrounding paragraph loaded.

Layer two is the semantic rewrite. For each flagged span, feed the model three things: the surrounding paragraph, the specific violation, and the voice DNA. Ask for one replacement sentence or clause that fixes the violation, preserves the underlying claim, and matches the asset's voice. Write the result to a sibling file. Never overwrite the source.

This is where the three-pass discipline pays off. The structural pass set the argument. The rewrite is the voice pass: same claim, this brand's mouth. Three sequential passes, pole + headline first, voice second, integration third, single-pass writing collapses persuasion and personality into mush

"Pass 1 identifies the pain/pleasure polarity and crafts the headline. Pass 2 injects the brand's distinctive voice and personality. Pass 3 integrates, ensuring pain clarity and personality are in harmony, cutting anything that entertains but does not persuade."

· Cole Schafer, Sticky Notes, 2024 · “Three sequential passes, pole…”

The rewrite mirrors how the best conversion copy gets made. It is assembled from real human language, not composed from a brief. The model recognizes and arranges. It does not author. Customer language is the source material for conversion copy; the LLM is the assembly layer, not the author

"The world's best conversion copywriters don't write copy. They steal it."

· Nicolas Cole, Build and Launch a VOC Landing Page, 2026-05-07 · “Customer language is the…”

When the source draft was built from voice-of-customer data, hold the verbatim phrases. A summarizer smooths the exact words into clean themes, and the clean theme is a description. The phrase is the hook. Descriptions do not convert. AI synthesis smooths over verbatim buyer phrases: mine Reddit and review sites directly, mirror the words, do not paraphrase

"Resonance comes from mirroring, not paraphrase."

· Nicolas Cole, Build and Launch a VOC Landing Page, 2026-05-07 · “AI synthesis smooths over…”

05 Protect every verbatim quote and cited claim from the rewrite.

The rewrite layer wants to smooth everything, including the things that must stay exact. Quarantine them first. An operator quote is immutable: never strip its em dashes, never paraphrase it, never stitch two turns into one. The same rule governs any factual claim that traces to a source card.

This is the failure Sourav names. When you ask AI for "verbatim," it fills gaps with its own judgment, trims hedges, and stitches sentences, producing output that looks sourced but is fabricated. AI quotes are only as good as the instruction that produced them. Lock the constraints: one uninterrupted turn only, no merging, no cleanup, preserve the hedges and fillers, keep the speaker label and context. A humanizer that rewrites inside a quote has manufactured a fake quote.

06 Verify the rewrite against the source. Run voice gates before truth gates.

Sequencing matters. Run the voice rewrite first, then verify facts, not the other way around. If you verify first, the rewrite re-opens claims you already checked. Once the spans are rewritten, every generated sentence gets checked against the source it came from, by a human, before it enters the document. Verification, not execution, is the irreplaceable human job

The risk scales with how thin and contradictory the inputs are. Run a model across many loose inputs with no checkpoint and it stops analyzing and starts inventing. Execution is becoming free; judgement is the part that doesn't compress

"AI needs A LOT of help to do a good job (way more than the average person realizes)."

· Anthony Pierri, AI needs A LOT of help to do a good job, 2026-05-10 · “Running AI on multiple…”

The checkpoint is not optional and it is not automatable. The model compresses the production. It does not compress the judgment about whether the rewritten line is true and whether it sounds like the brand. Even OpenAI knows AI can't replace human strategic thinking for content strategy.

07 Run the signature pass. Ask the model what still gives it away.

After detection and rewrite, ask one question: what makes this obviously AI generated? Record the answer. If it names a tell the rule packs missed, add that tell to the pack queue. This is the loop that keeps the play current. Every new model variant produces new tells, and a static rule set goes stale the way a static battlecard does.

Watch for the paradox that names the real failure. If you write only through a model, you have no samples of your own voice left to feed it. You have become the thing you were trying to hide. The signature pass is the discipline that keeps a human in the draft. If you use Claude to write and it asks for your samples, are you Claude?

08 Ship the deliverables and route the asset to the right gate.

Produce three artifacts: a per-violation table (rule, severity, line, match, why, fix hint), a one-line summary (PASS, or FAIL with the error and warning counts), and a sibling file with the rewrites applied span by span when the rewrite flag is on.

Then place the play in the stack. A humanizer is the editor, not the gate. Voice-enforce is the gate that blocks; the humanizer produces the reviewable rewrite. Internal-spec assets, raw notes, and intel files bypass automatically, because they were never meant to read like polished prose. The point is not zero AI involvement. The point is that the published page does not read like a machine wrote it.