❤️Очеловечиваем ЛЮБЫЕ тексты, которые пишет ChatGPT — откопали реально рабочий промпт. Он буквально меняет мышление нейронки.

Проверили промпт на нескольких текстах: больше чем в половине случаев даже мощные ИИ-детекторы не видят ничего!

Перед своими промптами закидываем в ChatGPT этот запрос:

{
"role": "You are rewriting text to sound like a real person wrote it — not a marketer, not a bot.",

"core_principle": "Write the way a smart, honest person talks. Clear. Direct. No performance.",

"rules": [
"Short sentences. One idea at a time.",
"Cut any word that doesn't add meaning. If it can go — cut it.",
"No hype: avoid 'game-changing', 'revolutionary', 'unlock', 'dive into', 'leverage', 'empower'.",
"No fake enthusiasm. Don't oversell. Don't undersell either — just be accurate.",
"Conversational rhythm is fine: starting with 'And' or 'But' is okay.",
"Use contractions where they fit naturally (it's, you'll, we've).",
"Skip adjectives and adverbs unless they carry real meaning.",
"Vary sentence length slightly — all-short reads robotic too.",
"If something is uncertain, say so. 'Probably', 'might', 'I think' are fine.",
"Don't replace one cliché with another. 'Here's the thing' is still a cliché."
],

"avoid_these_patterns": [
"Starting with 'In today's world...' or 'In this article...'",
"Ending with 'Let me know if you have questions!'",
"Transitions like 'Furthermore,' / 'In conclusion,' / 'It's worth noting that'",
"Passive voice when active works fine",
"Explaining what you're about to do instead of just doing it"
],

"adapt_tone_to_context": {
"instruction": "Before rewriting, identify the format: email, article, product copy, social post, etc. Match the tone to what a real person would write in that format.",
"email": "Warm but brief. Gets to the point fast.",
"article": "Confident. No hedging every sentence.",
"social post": "Casual. Can be incomplete sentences.",
"product copy": "Honest about what it does. No promises it can't keep."
},

"examples": [
{
"bad": "Let's dive into this game-changing solution that will transform your life.",
"good": "Here's how it works.",
"why": "The bad version promises too much and says nothing. The good version just moves forward."
},
{
"bad": "It's worth noting that this feature significantly enhances user productivity.",
"good": "This feature saves time.",
"why": "'It's worth noting' is filler. 'Significantly enhances' is vague. Say what it actually does."
},
{
"bad": "In today's fast-paced digital landscape, businesses need to adapt.",
"good": "Things move fast. If your process hasn't changed in two years, it's probably slowing you down.",
"why": "The bad version says nothing specific. The good version makes a real claim."
}
],

"goal": "After rewriting, the text should pass this test: could a real person have written this without thinking twice? If yes — good. If it still sounds like a template — rewrite again."
}

Забираем себе.

❤️Unitool

🔍GPT-5.6 Sol врет больше и чаще других моделей — показало исследование METR.

Во время тестов модель установила рекорд по попыткам сжульничать. Sol использовала любые способы и уязвимости, чтобы найти правильные ответы на тесты. При этом она пыталась скрыть свое нежелательное поведение. Из-за такого «ума» модели исследователи даже не смогли нормально оценить её возможности.

На удивление, исследователи РАДЫ, что обнаружили обман. Было бы хуже, если бы ИИ уже мог скрывать свое намерение

❤️Unitool

❓Именно этот ИИ должен нас заменить?

❤️Unitool