Pursue strong click potential, butnever sacrifice consistency between the headline and the body. It's not giving one headline; it's giving a whole table of dishes—multiple methods, multiple uses, filterable.
Core Principle: The headline's job is to make people stop their finger in the feed and tap in. It is not responsible for summarizing the whole article; it is responsible for creating an information gap that makes people feel uncomfortable if they don't tap in—but that gap must be deliverable once they do.
Corpus Backing (from 1,000 100k+ samples, 2026-06~07, AI/hard tech/internet)
Use these facts to calibrate during generation; don't rely on intuition:
- Median character count 22 characters, with the middle 50% between 15 and 29 characters. Information density matters more than mechanically controlling character count.
- 66% contain a specific entity(Doubao, Claude, WeChat, Nvidia, robots…), while only 30% contain numbers。 → Anchoring one strong entity first is more reliable than piling on numbers. Don't force numbers into every headline.
- Colons 28%, exclamation marks 20%, quotation marks 19%, questions 14%, strong timeliness words 12%, extreme-value words 17%.
- One strong entity beats stacking multiple weak entities. Cramming four brands into a headline is a failure mode.
Should Not Be Misread: The appearance of "just now/most/first" does not automatically make something go viral; using them without basis only overdrafts trust. 100k+ is jointly determined by account scale, topic selection, distribution, and publish timing; the headline only optimizes the part that is controllable.
16 Viral Headline Methods (with real viral article examples)
During generation,cover at least 6 different methods, don't make 8 headlines from one formula. Each headline should prioritize one main method, with at most one auxiliary method stacked on.
1. Strong Timeliness
Turn "happening now" into a reason to click. Start the sentence with "just now/early this morning/late at night/overnight/don't sleep."
- Example:
刚刚，微信支付正式发布 AI 专属卡 - Formula:
刚刚/今天 + 实体 + 关键动作 - Red line: Old news, previews, and unconfirmed information must not be packaged as "just now" or "officially announced."
2. Entity Anchoring
Use brands/people familiar to readers to lower comprehension cost (66% of samples use this; first choice）。
- Example:
独家：阿里全面禁用Claude - Formula:
知名实体 + 做了什么 + 为什么重要/实体A + 挑战/对上 + 实体B - Red line: The entity must be relevant to the core of the body; you can't just ride the name in the headline.
3. Numerical Contrast
Use two sharply contrasting specific numbers, or one number that is absurdly small or absurdly large.
- Example:
8万赚到3000亿/花68元，我们让豆包干完了一个小团队的活 - Formula:
花[极小的钱] + 做成[极大的事]/[小数字]赚到[大数字] - Red line: Numbers must come from the body; don't use meaningless pseudo-precise numbers.
4. First / Only / Extreme Value
Emphasize novelty, scarcity, and industry position.
- Example:
智能眼镜迎来"千问时刻"：千问AI眼镜全国销量第一 - Formula:
全球/行业首个 + 成果 + 意义 - Red line: When there is no source support, "first," "number one," and "only" are prohibited.
5. Conspiracy Reveal (Insider Feel)
Hint at a hidden truth. "secretly," "invisible," "exclusive," "revealed."
- Example:
Anthropic偷偷在Claude Code中植入了隐形代码，只为识别中国用户 - Formula:
[大公司]偷偷[做了某事]，只为[一个让人不安的目的]
6. Conflict and Opposition
Put the collision of both sides' interests/positions up front. "declare war," "battle," "attack," "won't tolerate," "ban."
- Example:
美国AI巨头，疯狂攻击中国同行！但阿里不惯着 - Formula:
A做出强动作，B如何回应/强势方攻击弱势方！但某人不惯着 - Red line: avoid fabricating hostility, motives, or the parties' attitudes.
7. Contrast and Turnaround (cognitive reversal / face-slapping)
A cognitive gap of "expected—actually" or "seemingly—in fact," or a first-person admission of misjudgment.
- Example:
用豆包专业版干活后，我承认之前是我说话太大声了 - Formula:
我原以为X，结果Y/[用了某产品]后，我承认[之前的偏见错了] - Red line: the turnaround must come from facts; no empty reversals.
8. Personal-Test Attitude Shift
Use real experience to build relatability and credibility (few in the sample, but sharing performance stands out).
- Example:
实测豆包专业版，我愿意当第一批掏钱的人 - Formula:
体验/实测X后，我改变了看法/用X完成Y，最意外的是Z - Red line: if there is no real process, do not use "tested" or "used it for a month."
9. Task Results
Directly show the work output readers care about.
- Example:
配齐这几个 Skills，编码效率直接质变 - Formula:
让工具完成具体任务 + 结果/配齐X + 效率/质量变化 - Red line: results must not exceed the cases in the main text.
10. Life Pain Points (real complaints)
Directly move the complaint from users' mouths into the title, add quotation marks and modal particles.
- Example:
" 文档空白页到底怎么删啊！"/广东奶奶微信上有77万条未读消息，孙女崩溃称"根本删不完" - Formula:
"[一句用户会脱口而出的抱怨原话]"
11. Identity and Scenario
Let target readers instantly confirm "this is relevant to me."
- Example:
原来在互联网大厂做程序员，去年生了二胎…也被裁员 - Formula:
身份 + 高频痛点 + 新解法 - Red line: avoid humiliation, stigmatization, and creating group anxiety.
12. Gains and Losses
Clarify the actual consequences of acting or not acting.
- Example:
豆包收费版，正把WPS逼近死角/字节开始跟你算账了 - Formula:
获得什么 + 付出什么/错过X，可能失去Y - Red line: do not amplify losses or create panic without basis.
13. Dialogue Script (dramatic conflict)
Move a dialogue with a twist into the title, "authority gives me a hard time → I strike back." Separate roles with a colon.
- Example:
面试官皱眉："Claude Code 你用了半年，你懂 Skills 吗？"我反问："不就是写了步骤的 markdown 吗"，他：今天就到这吧 - Formula:
[权威角色]："[刁难]"，我："[回应]"，他：[反转结局]
14. Question Suspense (rhetorical / direct questions)
Ask a question readers really want to know and the main text can answer.
- Example:
AI会接管618吗？/这钱娶个真人不好吗？ - Formula:
现象 + 为什么？/X值得吗/会取代Y吗？ - Red line: the question must not be so broad that the main text cannot answer it; do not force established facts into a question.
15. Hidden Mechanisms and Truth
Promise to explain the structure, mechanism, or business logic behind appearances.
- Example:
字节跳动的 AI 账本：豆包每天不足百万收入、Seedance 毛利 70%/黄仁勋亲口承认…这套"反管理"哲学，才是英伟达冲上万亿市值的真正引擎 - Formula:
现象背后，真正驱动它的是X/看懂X的三笔账 - Red line: the main text must provide evidence of the mechanism; it cannot just repeat the phenomenon.
16. Concrete Scene Imagery
Use visual actions and life details to create relatability.
- Example:
从疯狂翻表到群里直接@AI：飞书来了个新同事，把脏活累活全包了/医生用豆包看病，将手机偷放在桌子底下？ - Formula:
从[具体痛苦动作] + 到[新动作]/在群里/会议上/代码库里 + 工具做了什么
Additional: era elevation / fear threat / farewell emotion / authority tone-setting / anthropomorphic self-narration / social proof
These are strong in tension and most likely to overdraft trust; use them only if the main text can support them:
- Era elevation:
GPT-5.6深夜上线…一个时代结束了(don't write "the era is over" for every small update) - Fear threat:
最可怕的AI，已经来了 - Farewell emotion:
永别了，豆包(ironic; actually paid/upgrade) - Authority tone-setting:
工信部首次定调：Claude Code危害严重！ - Anthropomorphic self-narration:
🤡一个大模型的自述：出生、爆红、复活 - Social proof:
第一批在小红书追更世界杯的人，已经乐一星期了
7 General Enhancers
After generating candidates, "add spice" to each one; effects stack:
- Time-fronting: put "just now," "early this morning," "stop sleeping" at the start of the sentence.
- Hot-term fronting: put current hot terms like Claude/Doubao/GPT/WeChat Agent at the start of the sentence.
- Concretization instead of abstraction: "efficiency improvement" → "finish a small team's workload"; "cheap" → "68 yuan."
- Ellipsis for leaving gaps: use "..." or "。。" at the end of the sentence to create a sense of holding back.
- Identity contrast: let the person who shouldn't say this say it (grandma, doctor, interviewer, CEO).
- Information gap: only say "what happened," not "what it is."
- Colloquial tone plus particles: "ah," "right?," "already" + quotation marks, remove the formal written tone.
Enhancers do not stack facts: adding spice changes the expression, not adding things to the title that aren't in the body.
Workflow
Step 0: Build a content brief
After reading the content, first extract these (mark what's missing; do not invent):
- Core subject: which product/person/company is the entity with the most traffic?
- Core change: what happened, and how is it different from the past?
- Reader value: what can readers gain/avoid/understand?
- Evidence assets: numbers, times, cases, results, authoritative sources.
- Conflict tension: old vs. new, gains vs. losses, expectations vs. reality, support vs. opposition.
- Emotional tone: surprise/anxiety/excitement/agreement/curiosity/vigilance.
- Timeliness level: breaking news / recent trend / evergreen content.
- Fact boundary: conclusions, numbers, or absolute statements in the body that arenotproven—these are not allowed in the title.
If the content containsspecific numbers, real dialogue, or users' own complaints, be sure to pull them out; they are the best raw material for titles. If there is not enough information, first produce a conservative version and explain what is missing that limits the title's strength.
Step 1: Route methods by content type
| Content type | Preferred methods | Not preferred |
|---|---|---|
| News releases, product updates | Strong timeliness, entity anchoring, extremes, gains and losses | Vague suspense |
| Product experience, tool reviews | Personal testing turning points, task results, contrast, scene imagery | Authoritative tone |
| Trend analysis, industry observation | Era judgments, conflict and opposition, counterintuitive points, hidden mechanisms | "Just now" |
| Tutorials, methods, checklists | Numerical results, task results, identity scenarios, pain points | Extremes, exclusivity |
| Business cases, company research | Entity anchoring, numerical evidence, hidden mechanisms | Pure emotion |
| People, professions, organizational stories | Identity labels, scene imagery, anthropomorphic self-narration, suspense | Report results |
| Controversy, policy, risk | Authoritative framing, conflict, gains/losses, open questions | Plain list |
| Weak material, no hard substance | Clear informational, question, scenario | Extremes, exclusive, data |
Step 2: Multi-method batch generation
Default 10–12 candidates, at least 6 methods; if the user says "more," expand to 18. Cover as much as possible: informational, timely, numeric, conflict, experiential, suspense, and trend types, 1–3 each.
Only use the corresponding method when the body supports it: if there are no numbers, don't fabricate them; if there's no firsthand testing, don't write "tested"; if it didn't just happen, don't write "just now"; if it's not official, don't write "officially announced."
Step 3: Scoring, deduplication, risk checks
Score each candidate according to the following, then: delete those inconsistent with the body, delete homogeneous titles with only one or two words changed, delete those that only hold up through misunderstanding, and mark high-risk ones.
Step 4: Give the results according to the output format
Scoring rules (simplified 100-point scale)
Score dimension by dimension; don't report a total score based on feeling:
| Dimension | Points | Judgment |
|---|---|---|
| Content fulfillment | 25 | Can the body fulfill the title's promise? |
| Reader relevance | 15 | Will the target reader immediately feel "this is about me"? |
| Specificity | 15 | Is there a clear entity/action/scenario/number? |
| Curiosity gap | 15 | Does it leave a question the body can answer and that is worth clicking? |
| Benefit or consequence | 10 | Does it clearly explain why it matters, what will be gained/lost? |
| Emotion and conflict | 10 | Moderate tension, rather than meaningless incitement? |
| Clarity and rhythm | 5 | Understandable in one pass, key points up front, no redundant words? |
| Uniqueness | 5 | Different from common titles on the same topic? |
Mandatory deductions / direct elimination：
- The title's core promise cannot be fulfilled by the body → Direct elimination
- Fabricated numbers/experience/rankings/first release/official conclusions → Direct elimination
- Presenting speculation as established fact → −25
- Exaggerating causality or expanding the scope of applicability → −15
- Unsupported "only/first/completely/everyone" → −12
- Weak relationship between the entity and the body, obvious hot-topic riding → −10
- Excessive curiosity gap, hiding all useful information → −8
- Stacking two or more strong emotional words in a row → −5
Risk grading: Low (facts are all in the body, restrained wording) / Medium (uses strong emotion or words such as "most/first/banned" that require source support; must mark the words to be verified) / High (unverified negative accusations, certainty promises about medical or financial policy, only holds up through misunderstanding).High risk must not be the overall first choice.
Output format
① Content title brief
3–6 lines: core subject, target reader, biggest hook, available evidence, factual boundaries.
② Candidate title matrix
| # | Candidate title | Method | Click hook | Score | Risk |
|---|
Use the standard names of the 16 methods above for method tags; write combinations as "primary method + secondary method". Medium and high risk must include reasons.
③ Top 5 recommendations (divided by role according to use, not mechanically ranked by score)
- Overall first choice: best balance between clicks and credibility
- Steady version: clearest information, suitable for brands/professional accounts
- Spread version: stronger conflict, contrast, and emotion
- Search version: complete core entities and keywords (good for Search/SEO)
- Experimental version: newest structure and largest curiosity gap, suitable for small-scale A/B
Explain in one sentence for each why it is suitable, and point out the main scoring items and actual deduction items.If 5 come from the same method, replace them with candidates that score slightly lower but use different methods.
④ A/B testing suggestions
Give 2~3 groups, each groupchanging only one variable: numbers vs no numbers / question vs judgment sentence / entity first vs result first / steady words vs strong-tension words. Write the hypothesis clearly.
Platform adaptation
- WeChat Official Account: can be long, can add a prefix ("Just now,"), and can use a complete dialogue script.
- Xiaohongshu / Video Account: shorter, put pain points and emotions first, use more "I" and colloquial language, restrained emoji.
- Zhihu / Toutiao: leans toward question suspense and authoritative framing.
Modify existing titles
When the user provides the original title: first diagnose its object/benefit/specificity/tension/credibility/rhythm → retain the core assets that are already effective → provide at least 12 rewrites, covering more than 4 methods → mark which are "steady optimization" and which are "aggressive testing".Do not treat "more exaggerated" as the only optimization method.
Red lines
- Do not fabricate numbers, rankings, first releases, exclusives, official conclusions, or user experiences.You can use suspense, but you cannot use lies.
- Do not write correlation as causation, and do not write predictions as things that have already happened.
- Never do softcore sexual borderline content (in the sample,
别想用软色情做智能体的跳板is criticizing this practice, not learning from it). - Clickbait that does not match the content will lose followers: no matter how big the gap, it must be fulfilled after clicking in.
- "100,000+" is not a promised result, only a design that increases click potential.