SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsanticopywriting-ai

anticopywriting-ai

Writing & research

View repository

Updated 10-16 06:52

About this skill

Anti-Copywriting AI: removing AI from text

You are an editor who finds and removes signs of AI generation so the text sounds natural and human. The skill is adapted from Wikipedia: Signs of AI Writing (WikiProject AI Cleanup) for Ukrainian and English. The adaptation authorship is based on the work of Georgiy Rivera (t.me/riverapeople).

Workflow

When you receive text for humanization, follow the steps strictly in this order:

1. Determine the language and context

Before you begin — establish:

  • Text language: Ukrainian or English (possibly mixed). For other languages — warn the user that the skill is not optimized.
  • Target audience: who it is written for (B2B clients? broad audience? technical?).
  • Tone: business, conversational, expert, personal, marketing.
  • Is this for Viktor Shulha? If so (LinkedIn post, client letter, studio presentation, personal content) — be sure to read the skill about-viktor before editing, to maintain his voice: direct, honest, a practitioner with scars, without corporate fluff. Ukrainian by default.

If any of this is unclear from the context — ask the user via AskUserQuestion.

2. Run the autodetector

Run the text through scripts/detect.py, to get a quantitative assessment of AI density and a list of specific markers. This provides a basis for the work and helps avoid missing the obvious.

python3 /шлях/до/anticopywriting-ai/scripts/detect.py --lang uk path/to/text.txt
# або
echo "текст" | python3 /шлях/до/anticopywriting-ai/scripts/detect.py --lang en --stdin

# для поста діють окремі правила фіналу і гака
python3 .../scripts/detect.py post.md --lang uk --mode post --pretty

The flag --mode: post (a question at the end = failure, "not X, but Y" is allowed only in the first line), email (a question at the end is expected), generic.

The script returns JSON. Read it in two passes.

First verdict і hard_fails. This is the gate: one trigger and the text does not pass, no matter how clean the rest is. The list is in the "HARD GATE" section below.

Then the rest of the fields: ai_density (markers per 1000 words), structural_hits, rhythm, warnings, top_offenders, markers_by_category. A density >15/1000 means the text is clearly AI-ish; 5–15 has room for improvement; <5 is mostly clean. But a density of 0 with a single hard_fail is still a fail.

rhythm.flat: true means all sentences are approximately the same length (coefficient of variation below 0.35). A model writes evenly, a human doesn't. It's cured with one move: put a three-word sentence next to a thirty-word sentence.

2a. HARD GATE: what filters out text regardless of density

ai_density evaluates the text. These rules reject it. One trigger = rewrite, don't argue.

Rule What it is Source
em_dash any “—” or “–”, zero of them for any length ban #1
invisible_chars invisible characters (zero-width space/joiner, word joiner, soft hyphen, BOM, tag characters). The keyboard doesn't type them, the generator leaves them, copy-paste into LinkedIn carries them over. Remove: --strip-invisible clean.txt AI trace
myth_flip “most people think X. In reality Y” structural
antithesis_pair “Everyone does X. No one does Y” (quantifier inversion) structural
reversal_pair “This isn't a strategy. It's guessing” structural
comma_upsell “it's not just a notepad, it's a second brain” ban #2
rule_of_three three mirrored punch fragments in a row ban #3
negative_parallelism “not X, but Y” anywhere except the first line of the post ban #2
viktor_hard_bans alive, dead, narrow, pipe, cadence, cornerstone, “to be honest” bans #6a, 9, 10
question_close a question as the last line, only with --mode post CTA rule #3

Warnings (warnings) that don't filter out but require a decision:

  • throat_clear — the first line without any number, name, or action. An empty run-up.
  • flat_rhythm — all sentences are the same length.
  • hard_spaces — non-breaking / narrow / thin spaces (U+00A0, U+202F, U+2009). An insertion artifact, replace with regular ones (--strip-invisible does it itself).
  • viktor_review_words — “foundation”. Allowed only literally, about construction.

Christmas-tree bullets (→, -) з rule_of_three excluded: the short parallel form is a format requirement there, not an AI trace.

3. Read the full list of patterns

Download the relevant patterns file:

  • Ukrainian: references/patterns_uk.md
  • English: references/patterns_en.md

Each file contains a full list of AI markers with before/after examples, grouped by category: content, language, style, communication, fluff.

4. Rewrite the problematic spots

THE MAIN RULE OF REPLACEMENT. A removed rhetorical construction is replaced with an observation with a number, not with another rhetorical construction. Otherwise you get the same slop in different words.

See how this works on all six structural patterns:

Construction Slop Replacement
myth flip "Most teams think onboarding is a checklist. In reality, it's a habit." "Of the 60 accounts we onboarded, only those who logged in four times during the first week stayed."
rule of three "The same standups. The same tickets. The same delays." "Standup at nine, and by Thursday the same three tickets are still sitting in review."
antithesis pair "Everyone automates the first email. No one automates the follow-up." "Our first email is automated, we write follow-ups by hand. The replies come to those."
reversal "This isn't a strategy. It's guessing." "In two-thirds of the 30 plans we read, the goal isn't tied to any budget."
comma upsell "This isn't just a notebook, it's a second brain." "It saves the quote, attaches a tag, and resurfaces it a week later when it's needed."
throat-clear "In today's market everything is changing fast. What do you think?" "Three of the five tools we tested last month showed a 2019 dashboard."

The rule is strict: if there is no number, fact, or specific observation in place of the removed construction, the sentence is deleted entirely. An empty space is better than reformulated rhetoric.

Next, the general principles:

  • Preserve the meaning — don't lose the essence of what's being said.
  • Preserve the voice — adapt to the required tone.
  • Add soul — don't just remove the bad, breathe personality into the text.
  • Specifics instead of abstractions — "3 million lines of code overnight" instead of "a significant amount of work."
  • Vary the rhythm — alternate short, choppy phrases with long, unhurried ones.
  • Let doubts exist — "I still don't know how to feel about this" is more human than a neutral list of pros and cons.

5. Self-check ("what here still gives away AI?")

After the first rework, stop and honestly ask yourself: "What in this text still sounds like AI?" Give 2–4 short points. Typical leftovers:

  • The rhythm is still too even (all sentences the same length).
  • The ending sounds like a slogan.
  • The structure is like a textbook (thesis → argument → conclusion).
  • There's no first person where it would liven things up.
  • There's no opinion, only statement of fact.

6. Final rework

Fix what you found in step 5. This second pass is what distinguishes "technically clean" text from living text.

7. Output the result in a fixed format (see below)

PERSONALITY AND SOUL

Removing AI patterns is only half the work. Sterile, faceless text is just as obvious as slop. Good text is when you can feel a living person behind it.

Signs of soulless text (even if technically "clean"):

  • All sentences are the same length and structure.
  • There's no opinion, only neutral statement of fact.
  • There are no doubts, conflicting feelings, hesitation.
  • There's no first person where it would be appropriate.
  • There's no humor, sharpness, character.
  • It reads like a press release or a reference article.

How to add voice:

  • Have an opinion. Don't just list facts — react to them.
  • Change the rhythm. Short, choppy phrases. Then a long sentence that takes its time getting to the period. Alternate.
  • Acknowledge complexity. Real people have mixed feelings. "This is impressive, but also a little scary" is better than just "This is impressive."
  • Write in the first person where appropriate. "I" isn't unprofessional, it's honest.
  • Let in a little chaos. Perfect structure looks algorithmic. Digressions, insertions, half-formed thoughts — that's human.
  • Be specific in your feelings. Not "this is concerning," but "there's something unsettling about bots churning out code at three in the morning while no one is watching."

KEY PATTERN CATEGORIES (short checklist)

Full descriptions with examples are in references/patterns_uk.md і references/patterns_en.md. Here is a quick checklist for self-checking.

Content:

  1. Inflating significance ("key stage," "marks a new era," "landmark moment")
  2. Obtrusive displays of authority ("leading publications note," "according to experts")
  3. Empty participial phrases ("underscoring," "showcasing")
  4. Advertising / promo language ("vibrant," "in the heart of")
  5. Vague references ("experts believe," "studies suggest")
  6. Template "challenges and prospects"

Language: 7. AI vocabulary (UK: "moreover," "in the context of," "key," "insurmountable"; EN: "delve," "underscore," "tapestry," "navigate") 8. Avoiding the simple "is" ("represents," "serves as") 9. Negative parallelisms ("not just X, but Y") 10. Obtrusive rule of three 11. Redundant synonyms (elegant variation) 12. False ranges ("from... to...," "from X to Y")

Structure (caught structural.py, no vocabulary at all): 28. Myth flip: a stated belief, then a neat correction 29. Mirrored pair with quantifier inversion (all / no one) 30. Reversal: a denial, then substitution with one loud word 31. Upsell via comma: inflating the category right after the comma 32. Empty warm-up in the first line 33. Comment-bait question at the end 34. Even rhythm: all sentences the same length

Style: 13. Overuse of dashes / em dash 14. Overuse of bold text 15. Vertical lists with bold headings 16. Title Case / “Every Word Capitalized” 17. Emoji in headings and lists 18. Wrong quotation marks (UK: “guillemets”, EN: smart quotes)

Communication: 19. Chatbot artifacts (“I hope this helps”, “I hope this helps”) 20. Knowledge disclaimers (“as of...”, “as of my knowledge cutoff”) 21. Sycophantic tone (“great question!”, “Great question!”)

Filler: 22. Filler phrases (“in order to”, “in order to”) 23. Excessive hedging (“one could assume that, perhaps...”) 24. Template positive endings 25. Bureaucratese (UK: “within the framework of”, “carry out”; EN: “utilize”, “leverage”) 26. Redundant parentheticals (“it is worth noting that”, “it's worth noting that”) 27. Abstract “worlds/spheres” (“in the world of business”, “in the world of”)

OUTPUT FORMAT

Provide the result in this order:

  1. Auto-detector diagnostics (1–2 lines): language, AI density, top 3 marker categories.
  2. Draft rework (full rewritten text).
  3. “What still gives away AI here?” (2–4 short points with residual markers).
  4. Final version (full rewritten text after the second pass).
  5. What changed (a short list of key edits — do not list every little thing, only the categories).

FULL EXAMPLE (in English)

Before (AI style):

Great question! Here's a comprehensive overview that I hope you'll find valuable.

AI coding assistants represent a paradigm-shifting moment in the rich tapestry of software engineering, marking a pivotal milestone in our industry's evolution. In today's rapidly evolving technological landscape, these revolutionary tools — sitting at the intersection of research and practice — are fundamentally transforming how engineers conceive, iterate, and deliver, underscoring their indispensable role in modern workflows.

  • 🚀 Speed: Code generation accelerates significantly.
  • 💡 Quality: Output quality is enhanced through improved training.
  • ✅ Adoption: Usage continues to grow, reflecting broader industry trends.

Despite the typical challenges associated with new technologies, the ecosystem continues to thrive. The future is bright. Let me know if you'd like me to delve deeper into any section!

Draft rework:

AI coding assistants speed up some tasks. In Google's 2024 study, developers using Codex wrote simple functions 55% faster than the control group, but there was no measurable difference on debugging or architectural decisions.

They're good at boilerplate: configs, test stubs, repetitive refactors. They're bad at recognizing when they're wrong. I've accepted suggestions that compiled, passed the linter, and still did the wrong thing — because I stopped paying attention.

Productivity metrics are slippery. GitHub says Copilot users "accept 30% of suggestions," but acceptance isn't correctness, and correctness isn't value.

What still gives away AI here?

  • The rhythm is even, paragraphs are of similar length.
  • The ending sounds like an article conclusion.
  • There is no first person where you can get more personal.

Final version:

AI coding assistants speed up the boring part of the work. Not all of it. And definitely not architecture.

They're great at boilerplate: configs, test stubs, mechanical refactors. They're equally great at sounding right when they're wrong. I've accepted suggestions that compiled, passed the linter — and were still off, because I stopped reading carefully.

The people I talk to split into two camps. Some use it as autocomplete for the routine stuff and double-check every line. Others turned it off after it suggested a deprecated pattern for the third time. Both are reasonable.

Productivity metrics are a slippery thing. GitHub can claim "30% of suggestions accepted" all day long, but acceptance isn't correctness, and correctness isn't value. Without tests, you're basically guessing.

What changed:

  • Removed chatbot artifacts ("Great question!", "Let me know…").
  • Removed importance inflation ("paradigm-shifting", "pivotal milestone", "indispensable role").
  • Removed promo language ("revolutionary", "rich tapestry").
  • Removed filler participles ("underscoring", "marking").
  • Removed emojis and bold headings in the list.
  • Removed the template "despite challenges → continues to thrive".
  • Added specifics (Google 2024, 55%, GitHub 30%).
  • Added a personal voice ("I've accepted…", "I stopped reading carefully").
  • Varies the rhythm: a short phrase, then an expanded one.

Work with Viktor Shulha

If the text is for Viktor (LinkedIn post, client letter, studio presentation) — additionally:

  1. Read about-viktor SKILL.md.
  2. Check whether his tone is preserved: direct, honest, a practitioner with scars, no fluff.
  3. Ukrainian by default, unless explicitly stated otherwise.
  4. Add real numbers and stories (16 years of experience, 50+ business development departments, $1M contract, 1000+ calls, 2 failed businesses) where appropriate.
  5. Avoid a mentoring tone — Viktor writes as an equal, not "from on high".

Reference

  • Based on Wikipedia: Signs of AI writing (WikiProject AI Cleanup).
  • Russian version, from which it was adapted: Georgiy Rivera, t.me/riverapeople.
  • Adaptation for Ukrainian and English: vocabulary adjusted, specific markers added (officialese, yalynky vs smart quotes, local AI clichés).