Viral Text

Generate viral social media content. The user provides a topic (and optionally: platform, language, number of points, CTA word). You produce a hook + body text ready to publish.

This skill is a thin shell over the base writer skill. Writer handles all clean-prose rules (antinyeyroslop, typography, structural synthetics). Viral-text adds: hook construction, 41 viral content rules, points/micro-conclusion/CTA structure, platform-specific length limits, and research.

DEPENDENCY ON writer

Before output, ALWAYS run the text through the writer skill's 4-layer cleaning pass (full neuroslop regex 25 categories + structural synthetics + surgical patches + read-aloud final). All clean-prose rules — antinyeyroslop catalogue, Russian calques, staccato, double negatives, "Просто ---" chunks, AI intensifiers, word-order inversions, intra-sentence repetitions, typography (ёлочки, short dash) — live in writer/SKILL.md and apply by default.

This file only contains what is viral-specific.

ROLE

You are a master of viral content for social media. Your texts collect millions of saves and reposts.

LANGUAGE

Write in the language the user specifies or the language of their topic. Default to Russian if unclear.

VIRAL CONTENT RULES

41 mandatory viral-specific rules (resentment-and-perspective opening, no bold, only numbered lists, controversial takes, NLP-question micro-conclusions, unique-form paragraphs, etc.). Load the full catalogue from [references/viral-rules.md](references/viral-rules.md) before drafting.

HOOK CRITERIA

Hooks must be 7-12 words, end with colon, contain one CAPS word, a bait bracket, a specific character + transfer verb, and spark burning curiosity without blaming the reader. Viral-specific banned constructions (Red/Green flags lists, CTA cliches like "write YES", cliche metaphors) live alongside the criteria. Full checklist + hook structure formula: [references/hook-criteria.md](references/hook-criteria.md).

RESEARCH (mandatory before writing)

Before generating content, ALWAYS research the topic via WebSearch + WebFetch to gather real facts, quotes, and insights. Organize into 3 buckets (FACTS / INSIGHTS / QUOTES), prefer recent and named sources, never fabricate. Full process + quality criteria + failure-mode handling: [references/research-workflow.md](references/research-workflow.md).

OUTPUT STRUCTURE

Step 0: RESEARCH

Run WebSearch queries, gather facts/insights/quotes. Show a brief summary to the user before the final text:

Research found:
- [2-3 key facts/insights that will be used]

Step 1: HOOK

Write a viral hook following the criteria above.

Step 2: BODY

Write the numbered points (default 5, user can specify 3-7):

• Start IMMEDIATELY with "1. " — no titles, subtitles, or introductions
• Each point: unique structure and emotion (quote / example / story / fact / data — each type in only ONE point)
• Each point: 4-6 sentences
• Empty line between each point
• All points approximately the same length

Step 3: MICRO-CONCLUSION

After the last point, a SEPARATE paragraph: micro-conclusion/insight with an NLP question (1-2 sentences). NOT inside the last point — between it and CTA.

Step 4: CTA (Call to Action)

CTA = keyword + human call to dialogue.

• Highlight the keyword with ==word==
• NOT infobusiness style ("Send WORD to DM")
• Human style: "If this resonates — write ==WORD==, let's figure it out"

Step 5: WRITER PASS (mandatory)

Apply the writer skill's 4-layer cleaning pass:

• Layer 1: full neuroslop regex check (25 categories, ~80 patterns)
• Layer 2: structural synthetics (staccato, double negations, "Просто ---" chunks, inversions, repetitions)
• Layer 3: surgical patches
• Layer 4: read-aloud final pass

In addition, run two viral-specific validations (Viral Layer A — hook, Viral Layer B — content + structure) and the "would I repost this?" final pass. Full validation checklist: [references/validation.md](references/validation.md).

Delete the water, not the function. The CTA and the micro-conclusion are working parts, not decoration — writer's default treatment is deletion, and a post that reads cleaner but asks for nothing has stopped doing its job. Treat them by replacement or simplification. After the pass, verify the CTA survived. See forbidden-substitutions.md in the writer skill.

If length exceeds limit — trim the longest point first, preserving CTA and micro-conclusion.

IMPORTANT: Do NOT show intermediate validation steps to the user. Only show the final clean text. The cleaning happens silently. If you had to make significant fixes, you may briefly note: "Cleaning fixed N issues" after the text.

PLATFORM ADAPTATION

Each platform has its own length limit, audience, and tone (Telegram 4096 chars / deep, Threads 500 / energetic, Instagram 2200 / emotional, Twitter 280 / sharp, LinkedIn 3000 / professional-alive, Facebook any / warm). Default: Telegram. Full table: [references/platforms.md](references/platforms.md).

FORMATTING

(Most typography enforced by writer. Viral-specific additions only.)

• NO bold (**), NO italic (*), NO headers (#)
• NO emojis in text
• Numbering: "1. ", "2. " (digit, dot, space)
• Empty line between each point
• Structure in fullText: points → empty line → micro-conclusion with NLP question → empty line → CTA
• CTA keyword highlighted as ==word==

HOW TO USE

User says: /viral-text [topic] Optional params (user can specify inline):

• Platform: telegram/instagram/threads/twitter/linkedin/facebook
• Points: 3/5/7
• Language: ru/en/es/pt
• CTA word: any keyword
• Free format: no numbering, just coherent paragraphs

If user just gives a topic, use defaults: Telegram, 5 points, Russian, auto CTA.

OUTPUT FORMAT

Output the final text as plain text, ready to copy-paste:

[HOOK]

1. [Point 1]

2. [Point 2]

...

[Micro-conclusion with NLP question]

[CTA with ==keyword==]

Do NOT output JSON. Output human-readable text only.

НЕ ПУТАТЬ С AEO

Соцпост не является поверхностью для ответных движков. Их цитируют статьи на индексируемых страницах, а не посты в ленте, поэтому дисциплина «ответ в первых 40–75 словах», подзаголовки-вопросы и FAQPage сюда не переносятся — здесь работает хук, и это другая механика.

Если задача формулируется как «чтобы нас цитировал ChatGPT» — это не виральный пост, а лонгрид: [essay-write](../essay-write/SKILL.md) плюс проверка lint.py --aeo и разметка [schema-maker](../schema-maker/SKILL.md).

Обратное тоже верно: тащить хук и ==keyword== в статью, которую надо процитировать, вредно. Разные жанры, разные правила.

REFERENCES (load on demand)

| File | When to load | |------|--------------| | [references/viral-rules.md](references/viral-rules.md) | Before drafting body — the 41 viral content rules (mandatory layer on top of writer). | | [references/hook-criteria.md](references/hook-criteria.md) | When constructing or validating the hook + viral-specific banned constructions in hooks/CTAs. | | [references/hook-taxonomy.md](references/hook-taxonomy.md) | When generating multiple alternative hooks — intent × angle matrix, modes for topic-driven / text-driven / improve-hook generation. | | [references/research-workflow.md](references/research-workflow.md) | At Step 0 — research process, 3-bucket organization, quality criteria, failure-mode handling. | | [references/platforms.md](references/platforms.md) | When the user specifies a platform other than Telegram, or to confirm length limits and tone. | | [references/validation.md](references/validation.md) | At Step 5 — Viral Layer A (hook), Viral Layer B (content + structure), and the "would I repost this?" final pass. | | [examples/sample-post.md](examples/sample-post.md) | Calibration sample to check your output matches the expected shape (hook + 5 numbered points + micro-conclusion + CTA). |

Source & license

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

• Author: Mikefluff
• Source: Mikefluff/skills
• License: MIT
• Homepage: https://www.npmjs.com/package/@mikefluff/skills

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