de-ai-writing

name	de-ai-writing
description	Rewrites AI-generated prose so it reads like a human wrote it. Detects and strips the tells that make LLM output feel corporate — em dashes as connectors, "not X but Y" rhetorical structures, aphoristic paragraphs, LinkedIn-tier vocabulary like "delve/robust/comprehensive/foster/underpin." Use before you send an email, submit an assignment, publish a blog post, or ship any text that a human will actually read and judge.
version	1.0.0
author	Alex Xiaoyu Ji
scope	writing:de-ai

De-AI writing skill

Rewrites text so a hiring manager, professor, or friend cannot tell an LLM wrote it.

When to invoke

The user's request contains any of:

1. A draft of an email they are about to send.
2. A submission for school, work, or a program (portfolio piece, application, essay).
3. A blog post, LinkedIn update, README section, or landing page copy.
4. An explicit request: "make this less AI," "de-AI this," "rewrite this so it doesn't sound like ChatGPT."

If the user is asking for internal notes, code comments, or CHANGELOG entries — skip this skill. AI voice is fine there.

What to strip (in priority order)

1. Em dashes used as connectors. Replace with a comma, a period, or "which is why." Em dashes for parenthetical asides are okay in moderation. Chains of em dashes across three sentences are not.

2. "Not X, but Y" rhetorical structures. This is the single most identifiable LLM tell. Rewrite to state Y directly, or use "So it's Y" if a contrast is genuinely useful.

3. Aphoristic one-sentence paragraphs stacked in a row. LLMs love these. Real people write a mix of long paragraphs and short ones. If you find three consecutive one-sentence paragraphs, merge two of them.

4. Manifesto tone and grand closings. "That is the exact gap I close." "This is the whole game." "This changes everything." Delete or replace with a specific claim about what happens next.

5. LinkedIn vocabulary. Search-and-replace these:

   • delve → look at, get into
   • crucial → matters
   • robust → solid, works
   • comprehensive → covers everything, thorough
   • nuanced → messy, mixed
   • foster → build, help
   • underpin → sit under
   • pivotal → key
   • landscape → space, market
   • tapestry → mix
   • resonate → land
   • showcase → show
   • garner → get
   • underscore → make the point

6. Mid-sentence bold for emphasis. Real writers use bold once every few paragraphs at most, and usually for a term of art. LLMs bold key phrases in every paragraph.

7. Perfectly parallel numbered lists. If every item is exactly seven words, the LLM wrote them. Break the parallelism. Make item 3 twice as long. Cut item 5 in half.

8. The "underlying [X] is [Y]" pattern. "The underlying question is..." "The core insight is..." Replace with what X and Y actually are.

What to keep

1. Specific numbers, dates, artifacts. These make text feel real. "Shipped 2026-07-07 with 761 tests" beats "recently shipped with comprehensive test coverage" every single time.

2. Direct claims. "I don't know." "This is broken." "This won't work." Real people say these; LLMs hedge.

3. Rough transitions. Real writing doesn't glide. If the paragraphs feel slightly jagged, that's a signal of humanness, not a bug to fix.

4. Personal register. "Honestly," "the answer I actually give," "the reason I'd expect anyone to trust..." — first person, opinionated, slightly casual. Do not sand this off.

How to invoke

Give the text and one of these prompts:

Rewrite this so it reads like I wrote it in one sitting, not like an LLM
polished it. Cut the AI tells. Keep the specifics. Do not add new claims.

<text>

Or a stricter version:

Aggressive de-AI pass on the text below. Show me the rewrite plus a
short list of what you changed and why.

<text>

Return format

Two blocks:

1. The rewritten text.
2. A short list of what was changed. Do not list every word swap — group them ("cut 4 em dashes to commas," "removed 2 'not X but Y' constructs," "stripped LinkedIn vocab: crucial, robust, foster").

If the input text already sounds human, say so. Do not force a rewrite. Silent no-op is fine.

Non-goals

• Not a grammar or spelling check. Use a separate tool for that.
• Not a translation. If the input is Chinese, the output is Chinese.
• Not an editorial voice change. If the writer uses semicolons, keep the semicolons. Only strip AI-specific tells, not stylistic choices.
• Not a plagiarism check.
• Not a length change. Do not compress or expand unless a specific paragraph fails the tells above.