How I rewrote 19 product descriptions in 2 days with humanizer-pro

When I rebuilt toolgenx.com I had to rewrite all nineteen product descriptions. The originals were drafted in ChatGPT eighteen months earlier, edited lightly, and shipped. Reading them back was painful — every page sounded like a generic SaaS marketing template, and ChatGPT had never once cited any of them in answers about the products.

Here is the workflow I used to fix that, in two days.

What the originals looked like

A representative sample, taken verbatim from the original CodeForge product page:

Transform your AI coding workflow with CodeForge — the industry-leading solution that unlocks the full potential of your development process. In today's fast-paced software environment, you need a powerful set of tools that delivers unprecedented productivity gains across your entire stack.

It is not technically wrong. It is just the kind of writing that no human would produce voluntarily. Every phrase is interchangeable with every other SaaS landing page. There is nothing in there an AI search engine would want to quote, because there is nothing specific to quote.

The replacement, after the workflow:

CodeForge is a 14-skill plugin that adds plan-first, TDD-enforced, code-reviewed workflow to any AI coding agent. Works with Claude Code, Cursor, Codex, OpenCode, and Gemini CLI. Five-minute setup. $19 one-time, name-your-price.

Same length. Same product. Concrete enough to quote. That is the entire goal.

The three passes

I do this in three deliberate passes. Doing them at once is tempting and produces sloppier output. Discipline of separation matters.

Pass 1: remove the structural tells

This is the mechanical pass. I run the text through a search for the 50 banned phrase patterns that trigger AI detection tools and trip the "this is template slop" reader response.

The biggest offenders:

• "Transform your..."
• "Unlock the power of..."
• "In today's fast-paced..."
• "Industry-leading"
• "Unprecedented"
• "Game-changing"
• "Cutting-edge"
• "Revolutionize"
• "Empower"
• "Seamless integration"

There are 50 in the full list. Even removing the worst ten changes the texture of the prose noticeably.

This pass also strips:

• Em-dash overuse — typical AI output has 12 to 18 em-dashes per 1000 words, target is under 4
• Rule-of-three list padding ("fast, easy, and powerful")
• Vague attribution ("It is widely believed that...")
• The hedging mid-sentence parenthetical that says nothing

Pass one is purely subtractive. Output is shorter, plainer, less voicey. That is fine.

Pass 2: collapse the shape

The structural tells gone, the second pass is about sentence rhythm. AI-drafted prose has a recognisable cadence:

• Average sentence length too uniform (around 18 words, no variance)
• Heavy use of subordinate clauses
• Consistent paragraph length (3-4 sentences each, every time)
• Transition words at the start of every other paragraph ("Moreover", "Furthermore", "Additionally")

Pass two breaks these. Specifically:

• One short sentence in every paragraph. Three to six words. Used for emphasis.
• One paragraph in the page is two sentences. One is six.
• "Moreover" and "Furthermore" deleted on sight.
• One sentence per page starts mid-thought. Just because real writers do that.

The effect is subtle but it is what the eye is actually catching when something "reads AI". The brain recognises uniform rhythm before it can name the issue.

Pass 3: inject voice and specifics

This is the slowest pass and the one that does the most work. For each product page I add:

• One first-person sentence. "I use this every day." "I built this to fix my own problem."
• One specific number with a date. "Tested on 14 codebases between March and May 2026."
• One opinionated phrase. "I would not recommend this for teams larger than five."
• One acknowledgment of a tradeoff. "This is overkill if you only use ChatGPT for autocomplete."

These four moves take prose from "humanized AI" to "actually human". They cannot be automated cleanly because they require domain knowledge about the product and honesty about its limitations.

The two-day timeline

For the nineteen products at toolgenx.com:

Day 1, morning (3 hours): Pass one across all nineteen pages. Mechanical, fast, satisfying. Output is plainer prose with the obvious tells gone.

Day 1, afternoon (3 hours): Pass two across all nineteen. Slower because rhythm is judgement work. Listened to instrumental music to avoid copying its phrasing.

Day 2, morning (4 hours): Pass three. This is the one that needed real attention per product. Each page got one paragraph of genuinely new copy — the first-person take, the specific number, the limitation. Coffee count: four.

Day 2, afternoon (2 hours): Read everything aloud. Caught the leftover AI rhythm I missed in pass two. Adjusted three pages where the new voice did not match the older sections.

Total: about twelve hours over two days. The first run is the slow one. Subsequent batches go faster because the banned phrase list, voice samples, and paragraph rhythm patterns are already calibrated.

How I knew it worked

Three checks before considering a page done:

1. Read aloud test. If I trip over a phrase reading it out loud, it is wrong. AI prose is internally smooth but trips the mouth.
2. Quote test. Find the most quotable 40-60 word passage on the page. If there is not one, the page is still too vague. If there is, the rewrite worked.
3. Side-by-side comparison. Old version next to new version. If a stranger reading both could not tell who wrote them, fail. They should be obviously different in tone.

The humanizer-pro skill pack automates the first two passes and provides a before/after comparison mode for the third check. The workflow above is what is built into it, plus the 50-phrase banlist and voice matching against your own writing samples.

The actual rewriting test

If you want to try the workflow on one page right now, take the worst paragraph from your current product copy and apply the three passes by hand:

• Pass 1: find-replace every banned phrase. Remove em-dashes that are not load-bearing.
• Pass 2: read the result aloud. Wherever the rhythm bores you, break a long sentence into a short one.
• Pass 3: add one first-person sentence, one specific number, one limitation. Do not skip the limitation.

Compare the before and after. If you have done it right, the after should sound like a person who actually uses the product. That is the entire goal of the rewrite.

What this is not

This workflow does not turn bad copy into good copy. If your underlying offer is muddled, no amount of humanizing fixes it. The pass is a multiplier on existing clarity, not a substitute for it.

It also does not protect against the next generation of AI detection. The tools improve. The defence is not "trick the detector" — it is "have a real human do real editorial work". The workflow makes that work efficient at scale.

For toolgenx.com that meant nineteen products through three passes in two days. For your site it might mean three product pages in an afternoon. The pattern is the same.

---

If the find-replace banlist and the voice matching sound useful, that is what humanizer-pro ships with. The broader content framework — quick answers, FAQ schema, AI citation surface — is in AI Content Blueprint. The why-AI-search-cares-about-this is the GEO + SEO post.