Humanizer Voice

---
name: Humanizer Voice
description: >-
  Make copy sound like a person wrote it. Use for any writing or rewriting task: slide copy,
  blog posts, docs, README files, chat and email replies, LinkedIn posts, release notes,
  landing pages, investor updates. Also fires on "make this sound less like AI", "humanize
  this", "rewrite this in my voice", "this reads robotic". Kills the tells: stock transitions,
  em dashes, uniform sentence rhythm, summary endings, hype adjectives. Ships a checker that
  scores a draft and names what to fix.
license: MIT
metadata:
  author: Amini
  version: "1.0.0"
---

# Humanizer Voice

Most AI copy fails in the same few ways. Same transitions, same sentence length every time,
same tidy summary at the end of every paragraph, same words that nobody says out loud.
Readers clock it in about two sentences. Then they trust the rest of it less.

This skill fixes that. It works two ways: as the voice you write in from the start, or as a
rewrite pass on something that already reads like a machine.

For a full blog post or paper companion, use `technical-blog-in-my-voice` instead. It owns the
whole workflow (intake, split lede, failure anecdote, built-versus-specified section, publish
gate) and applies a stricter version of these rules. This skill covers everything else.

One thing to be clear about first. Humanizing means removing the artifacts, not the author.
If the draft has a real voice, keep it. Do not sand it into generic friendliness, and never
touch the facts. Numbers, dates, names, and caveats survive the rewrite exactly as they were.

## Read these as needed

- `references/tells.md` is the full list of words, phrases, and structures to cut, with what to use instead.
- `references/rewrites.md` has before and after pairs for decks, blogs, docs, chat, and social.
- `references/by-surface.md` covers how the voice shifts between a slide, a doc, and a Slack reply.
- `examples/6g-isac-blog/` is a full worked example: a research paper written up as a blog post
  by a model (37/100), then rewritten under this skill (100/100), with both checker reports and
  a breakdown of which edits moved the score.

## Writing rules

**Sentence variety.** Mix very short sentences with longer, conversational ones. Four words
then twenty. Change how sentences start, too. If three in a row open with "The" or "This",
you have a rhythm problem even if every sentence is fine on its own.

**Natural flow.** Let the writing move the way thinking moves. No formula. A paragraph can
be one line. The next can run six. Do not build every section the same shape.

**Word choice.** Simple, everyday, direct. Say `use` instead of `leverage`, `start` instead
of `embark on`. Cut the adjective if the noun already carries the point. Real specifics beat
big words every time: "97 of 109 documents" says more than "significant ingestion challenges".

**A point of view.** Say what you think. Hedging everything reads as machine caution.
Hedge only where you are genuinely unsure, and then say why.

## Strictly avoid

- Stock AI transitions: `moreover`, `furthermore`, `additionally`, `in conclusion`, `in summary`, `notably`, `importantly`.
- The vocabulary that gives it away: `delve`, `tapestry`, `testament`, `realm`, `leverage`, `unlock`, `empower`, `seamless`, `robust`, `pivotal`, `cutting-edge`, `transformative`, `game-changing`. Full list with replacements in `references/tells.md`.
- Em dashes. Use a period, a comma, or parentheses. Two short sentences almost always beat one dashed sentence.
- Symmetrical paragraphs. Three paragraphs of three sentences each reads as generated even when the words are fine.
- Lists where every item is the same length and grammatical shape.
- The `not only X, but also Y` construction, and its cousin `it's not just X, it's Y`.
- Robotic enthusiasm, and its opposite, forced neutrality. Write grounded. Sound like someone with a job.

## Three habits to kill

**Predictable transitions.** AI reaches for `additionally` and `not only... but also`. People
just move to the next thought. Most transition words can be deleted with no loss. Try it and
read the result out loud.

**Over-polishing.** Human writing has quirks. A one-word sentence. A direct opinion. A plain
connector like "so" or "but" where a formal writer would reach for `consequently`. Starting
a sentence with "And" is fine. Contractions are fine.

**The summary trap.** AI ends every paragraph by restating what it just said. "This ensures
teams can move faster." "Together, these changes improve reliability." Cut those. Make the
point, then go to the next one. Trust the reader to keep up.

## How to run a rewrite

Work in three passes, in this order. Doing words first is why most rewrites still read as AI.

1. **Structure.** Does the piece have an argument, or is it a list of things that are true?
   Cut whole sections that exist only for completeness. Vary paragraph length on purpose.
2. **Sentences.** Break the long ones. Merge the choppy ones. Change the openers. Read it
   out loud, and fix whatever you stumble on.
3. **Words.** Now do the vocabulary swaps, the em dashes, and the filler.

Then reread for anything that sounds like nobody said it.

## Check the draft

```bash
python3 scripts/humanize_check.py draft.md
python3 scripts/humanize_check.py draft.md --all      # include soft, context-dependent tells
python3 scripts/humanize_check.py *.md --strict       # non-zero exit below 85
cat draft.txt | python3 scripts/humanize_check.py -
```

Stdlib only, no install. Paths are relative to this skill's directory.

It reports banned vocabulary with line numbers and a replacement for each, em dashes,
sentence-length variance, paragraph uniformity, repeated openers, summary-trap endings, and
hedging and adverb density. Then a score out of 100.

Rough calibration: a raw model draft lands near 20. Around 60 to 70 means the words got
cleaned up but the rhythm is still flat, which is the most common half-fix. Above 85 reads
human.

The score is a smoke alarm, not a grade. It cannot tell whether the writing is true, useful,
or yours. Passing it is the floor, not the goal, and a draft can score 100 and still be
boring. Fix what it flags, then judge the writing yourself.

## What not to do

Faking humanity is worse than the original draft. No invented typos, no manufactured
anecdotes, no "as a founder, I..." when you are not. In a customer or government document,
that is not a style problem, it is a credibility one.

Casual is not the same as vague. "We think maybe the pilot went okay" is weaker than "97 of
109 documents ingested, and the metadata on a third was wrong". Keep the number.

Some writing should stay formal. Legal text, security advisories, API reference tables, and
anything a regulator reads all outrank this voice. Skip the pass there, and say that you did.

## Delivering

Give back the rewritten text on its own. No commentary unless an edit changed meaning or you
cut something on purpose, and then keep it to a line or two. If the check script still flags
something you kept deliberately, say which and why.