De Ai Flavor

de-ai-flavor

Strip AI tone out of a piece of writing and rewrite it so it sounds like the author actually wrote it.

This skill does not invent voice. It needs a personal style profile to rewrite against. The profile lives in a Markdown file the user maintains. The skill reads the profile, then applies a fixed checklist.

When to use

• User hands over a draft and asks to remove AI tone
• User says: "polish this", "less AI please", "make it sound like me", "rewrite this in my voice"
• After generating any public-facing post / tweet / blog Markdown — run it automatically before finalizing

When NOT to use

• README files, technical docs, code comments — these can stay formal
• Drafts that don't have full logic yet (write first, de-AI later)
• Code itself

Setup: the style profile

Before first use, the user must create a personal style profile. Default expected location:

~/.claude/projects//memory/writing_style.md

If the file is missing, the skill should ask the user where their style profile lives, or help them bootstrap one from 3–5 of their best existing pieces.

A good style profile defines:

1. Narrative structure (does the author open with a scene? a number? a feeling?)
2. Sentence rhythm (long-short alternation? short sentences as standalone paragraphs?)
3. Tone (self-deprecating? hesitant? colloquial words to keep?)
4. Conclusion style (do they end on a quiet beat or a punchline?)
5. AI markers to ban (transition words, symmetric structures, table summaries, etc.)
6. Forced rewrite actions (concrete steps to apply on every pass)

An example profile lives in examples/style-profile-example.md.

Process

Step 0: Load the style profile

Open the profile file. Read it fully. Keep it in context for the whole pass.

Step 1: Diagnose — don't rewrite yet

Read the original draft once. Find every instance of these 6 AI-tone problems and list them for the user:

1. Transition glue: however, therefore, in conclusion, moreover, furthermore, worth noting that, what's more (然而、因此、综上、不仅、更重要的是)
2. Symmetric triples: "not A, but B" repeated, three-item parallel structures stacked
3. Punchline endings: every paragraph closes with a polished aphorism
4. Flat abstraction: the whole piece floats at concept level, never lands on a specific detail
5. Structuralist itch: 1/2/3/4 numbered lists, symmetric subheads, summary tables
6. Over-stuffed paragraphs: every sentence delivers value — no hesitation, no self-correction, no breathing room

Output format:

[AI-tone diagnosis]
1. Paragraph X — [problem type] — ""
2. ...

N issues found.

Step 2: Apply the 5 forced rewrite actions

These are the core of the skill. Apply each one:

• A. Open with a concrete scene. Not "I want to talk about X." Start from a moment — a date, a number, a feeling, a thing that happened.
• B. Break symmetry. Any time three or more items line up in parallel, kill at least one. Let the structure be uneven.
• C. Insert "I"-breaks in every paragraph. Each paragraph needs hesitation, self-deprecation, contradiction, or revision. If the paragraph reads clean, it reads fake.
• D. Don't end on a punchline. The best endings settle, they don't peak. Let the last line be quiet.
• E. Delete transition words. Grep the whole draft for the words listed in problem 1. Replace with paragraph breaks.

Step 3: Check abstract-concrete alternation

The deepest tell of AI writing is flat abstraction level. Walk through the rewrite paragraph by paragraph and tag each one "abstract" or "concrete":

• All-abstract paragraph → drop in a detail (number, scene, quote, physical action)
• All-concrete paragraph → pull one feeling up to an idea (but don't punchline it)

Target rhythm: abstract → concrete → abstract → concrete. Never two abstracts in a row.

Step 4: Output a before/after diff — don't overwrite yet

## What changed
(one paragraph: what was done, where the biggest changes are, why)

## Key edits

### Edit 1: Opening
Before: ...
After:  ...
Why:    ...

### Edit 2: ...
...

## Full rewrite
(the complete rewritten draft)

Step 5: Wait for confirmation before writing

Do not overwrite the source file until the user explicitly says "OK" / "apply" / "覆盖" / "confirm".

Anti-example

However, the rise of AI is reshaping content creation.
It's not a tool change — it's a paradigm shift.
We need to rethink three things:
1. The nature of content
2. The role of creators
3. The position of platforms
In conclusion, the future belongs to those who understand both AI and humanity.

Hits every problem: however / not A but B / numbered triple / in conclusion / aphorism close / 100% abstract.

Positive example

May 22 was my last day. I had dinner with old colleagues that week. People kept asking why I left. I thought about it for a while and the real answer turned out to be simple:

My sense of self got a little too strong.

It sounds kind of edgy when I say it out loud. But it's the most honest reason I have.

Concrete opening (date, dinner, the question), self-deprecation ("edgy"), short sentences, no symmetry.

Notes

• Emoji-numbered subheads (1️⃣ 2️⃣) are allowed if the user's style profile permits them. They're not inherently AI.
• Technical writing (code, API tables, model comparisons) can stay slightly more formal — but skeleton, rhythm, tone must still follow the profile.
• "Sounding human" does not mean rambling. Real ≠ verbose. Cut first, then check.
• Cutting words beats adding words. AI-heavy drafts usually need 20–30% removed to feel right.

Project layout

de-ai-flavor/
├── SKILL.md                          ← you are here
├── README.md                         ← Chinese intro
├── README_EN.md                      ← English intro
├── LICENSE
└── examples/
    ├── style-profile-example.md      ← what a personal style profile looks like
    └── before-after-example.md       ← a real rewrite, showing the full process

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: anneheartrecord
• Source: anneheartrecord/charles-skill
• License: MIT
• Homepage: https://www.charles-cheng.com/projects

Install and usage instructions live in the source repository linked above.