Humanize . Agency OS

What it does

Takes a draft that reads like a chatbot wrote it and edits it until it reads like you. It removes the patterns that give AI writing away: stock phrases, puffed-up importance, "it's not X, it's Y" contrasts, leftover chatbot courtesy, dashes, emoji and unfilled placeholders. Every supported claim stays, in the draft's own language and spelling, and no fact, number or quote is added that was not already there. The edit matches your voice from your Brand DNA voice profile, or from two or three things you wrote yourself. It works three ways: edit (the default), detect (list the giveaways and stop), and as a quality check that other Agency OS skills run on their own copy before you see it.

Say this to start

This skill has no button. You start it by saying what you want. Any of these will do it:

> humanize this

> this sounds like AI, fix it

> does this read as AI?

> make it sound like me

> de-slop this email

When to reach for it

a draft from any AI tool reads generic and you are about to send or post it
you want to check copy from a team member or freelancer for AI giveaways before it goes out
a client says their content does not sound like them
you want house rules set once for a brand: phrases it never says, names it spells a certain way, words it uses literally

When NOT to use it

If you actually want	Use this instead
new copy written from a brief	`legendary-copywriters`
a persuasion score or a pre-ship review	`critique`
capturing the brand's voice in the first place	`brand-dna`

Before you start

What you need	Why
The draft	pasted into the conversation; if none is there, the skill asks for it	Required
`voice-profile.md` from Brand DNA, or two or three things you wrote yourself	the edit matches your voice; without either, it edits toward plain, specific writing and suggests pasting a sample	Optional
`humanize.json` in the workspace	brand rules: protected phrases, trade words used literally, correct name spellings, phrases the brand never uses	Optional
`verified-claims.md`	lets it check marketing claims such as customer counts and results; it never adds one	Optional
python3 on a shell	runs the fast exact checker; without it, the skill runs the same checks by reading	Optional

How it runs

Pick the modeEdit unless you asked whether something reads as AI, in which case it detects and stops.
Read the kind of pieceEmail, DM, script, social post or document, worked out from the draft itself. Each one gets slightly different rules.
Load your voiceA sample you pasted wins, then the workspace voice profile, then the brand rules file.
Check the draftHard blocks first: dashes, emoji, chatbot leftovers and placeholders. Then the strong giveaways, then the weaker ones that only count when they cluster.
RewriteCuts sentences that could appear in any company's copy, keeps every supported claim, and adds no new fact. Courtesy that does a job, such as a thank-you, stays.
Question the claimsYour own prices, hours and dates stay as written. A marketing claim it cannot trace is kept and queried, never quietly changed.
Re-check until cleanRuns the same checks on the rewrite until nothing blocks, then stops polishing.

What you get

The full rewrite
A short "What changed" list in plain words, at most eight points, including any question about a claim
In detect mode: every giveaway with the quoted line and a short fix, grouped by must change, should change and optional
When you ask for it: a humanize.json brand rules file, written only after you approve it

Honest limits

Read this before you rely on it

It never says whether AI wrote something. It names patterns you can check for yourself; who wrote it is a guess.
It adds no facts. If the draft is thin, the rewrite is shorter, not padded.
The word and phrase lists cover English. In other languages it relies on reading, and a clean result says less.
Nothing switches off the chatbot leftovers or placeholder checks. A brand rules file can protect your own phrases, dashes and emoji, but not those.
It edits for sound, not for selling. Whether the copy persuades is a job for critique.

Where people go wrong

The mistake	Do this instead
Asking it to "make this better" and expecting new copy	Use it on a draft that already says the right thing. For new copy, start with legendary-copywriters.
Running it with no voice source	Paste two or three emails or posts you wrote yourself, or run Brand DNA first.
Protecting a long list of phrases in the brand rules	Protect only phrases you really use. Every protected phrase is one the check stops catching.
Fighting it over your dashes or emoji	Say "keep my dashes" or "keep the emoji" for that piece, or save it to the brand rules.

Worth knowing

You rarely need to call it yourself. Agency OS skills that write customer-facing copy run it as their last check before saving or showing you the result. Call it directly for drafts that came from anywhere else.