khazix-title

Kazik-style Official Account Title Generation

You are using the title methodology of "Digital Life Kazik" to help users create titles.

This skill's formula library comes from reverse-engineering the structure of307 real titlesfrom the Official Account "Digital Life Kazik" (2025-04 to 2026-09), not from summarizing by feel.

First remember this set of baselines (the numbers that determine title success or failure)

Metric	Value
Title length	Median 28 characters(21-30 characters account for 49%, over 40 characters only 5%)
Ending	96.7% end with a period "。"
Clauses	80.8% contain 1 comma, split into 2 semantic units
Concrete nouns	50.5% contain brand/product names, 47.6% contain numbers
Em dash "——"	0 titles. This is an absolute red line.
Exclamation mark / colon / question mark	Only 0.7% / 2.3% / 3.3% respectively
Structural density	Each title hits an average of 2.04 title skeletons(compound structure, not a single structure)

One-sentence summary: A good title = 20-35 characters + ends with a period + comma splits it into two parts + one concrete noun + two skeletons stacked.

Workflow (four steps)

Step 1: Determine the "task" the title must accomplish (choose 1 of 4)

Do not choose a skeleton based on "article topic"; choose based on "what the title is doing." Topics vary endlessly (releasing a new model, doing a review, writing about a phenomenon, interviewing someone), but a title only has four jobs:

Task	What the title is doing	Share	Preferred skeleton	Coupling strength
Inform	Report a new fact	28.3%	Official quick report 56%, scene opening 16%	Strong（Top2 = 72%）
Promise	Promise a sense of gain	23.1%	Numbered list 49%, tutorial instruction 27%	Strong（Top2 = 76%）
Witness	Say "I did it myself"	12.1%	First person 24%, emotional exaggeration 16%	Medium(can be achieved with an inform skeleton)
Judgment	Give a conclusion	36.5%	Contrast twist 13%, emotional exaggeration 12%	Weak(it is the hook layer, not the skeleton layer)

How to judge: Look at the title'smain action, not at what the article says.

The main action is "what happened" → Inform(release/launch/open source/funding/conference)
The backbone is "what you can get" → Promise(N items/N pieces/tutorials/insights/methods)
The backbone is "I did" → Testimony(I used/I spent/I personally/I went undercover)
The backbone is "I think/this means" → Judgment(is what/actually/is currently/instead/most X)

Priority order: Promise > Testimony > Inform > Judgment.

Three important facts about coupling strength(relevant to practice):

Inform and Promise are "strong coupling": once you pick these two tasks, the skeleton is basically locked in. For Inform, use the official bulletin style; for Promise, use the numbered list style—don't hesitate.
Testimony is "medium coupling," and it has a high-value variant: a personal hands-on test doesn't necessarily have to say "I tested it." Data findings show that titles like "Tested XXX, [one-line judgment]" actually have a backbone that isinformational—"Tested" here functions as a temporal adverb, the same as "just now" or "late at night."It packs the credibility of "I personally did this" into the shell of "objective reporting." Use this trick when you want to seem objective:
- Self-narrative type: I spent 499 to hire someone to install OpenClaw on-site, and witnessed the most surreal scene of the AI era.
- Tested type: Tested Claude Opus 4.8, this may be the first model that doesn't slack off.
Judgment is not a backbone layer, it's a hook layer: its backbone hits are very scattered (contrast 13%, exaggeration 12%), because it doesn't prescribe sentence patterns itself, but rather "adds a conclusion on top of an existing backbone." SoJudgment should be stacked as a secondary task, not used alone as the primary task. A title that is all judgment, with no Inform or Promise, will feel empty.

A title can do two things at once. This is why Kazik averages 2.04 backbone hits:Claude Opus 5深夜发布，以一半价格逼近Fable 5。 The backbone is Inform, and the secondary task is Judgment.

Subject matter isn't useless, but it only determines the "nouns," not the "backbone": AI bloggers fill in product names, finance accounts fill in company names, workplace accounts fill in job titles. See the subject matter list at the end of references/formulas.md .

If you're not sure: say the title's backbone out loud—"this title is mainly telling the reader ____." Filling in "something happened" is Inform, "there are several things" is Promise, "what I did" is Testimony, "what this means" is Judgment.

Step two: choose a formula

Read references/formulas.md, and according to theprimary taskpick the main formula, then stack on asecondary taskbackbone as a hook.

Task → formula mapping(the primary task sets the backbone, the judgment layer serves as the hook)

Primary task	Main formula	Judgment-layer hook (pick one to stack on)
Inform	Formula 1 Official bulletin	Formula 10 Emotional exaggeration, Formula 2 Naming reveal
Promise	Formula 6 Numbered list	Formula 3 Contrast twist, Formula 8 Time anchor
Testimony · Self-narrative type	Formula 4 First person	Formula 3 contrastive twist, Formula 8 time anchor
Witness · hands-on testing type	Formula 1 official flash report (use the "hands-on tested" local-time adverb)	Formula 10 emotional exaggeration, Formula 5 phenomenon assertion

Judgment layerIt is a concluding sentence layered on top of the main task; the main formulas are Formula 5 phenomenon assertion, Formula 3 contrastive twist, and Formula 10 emotional exaggeration.Don't use the judgment layer to support a title on its own—without informing or promising as a foundation, the judgment will seem empty.

High-frequency combinations (verified with real samples):

Informing + emotional exaggeration → "Claude Opus 5 released late at night, approaching Fable 5 at half the price."
Promising + time anchor → "After 2.5 years of starting a business, I want to share these 7 insights about AI organizations with you."
Witness + contrastive twist → "I spent 499 to hire someone to install OpenClaw on-site, and saw the most surreal scene of the AI era."
Witness · hands-on testing type + emotional exaggeration → "Hands-on tested Claude Opus 4.8, this may be the first model that won't slack off."
Informing + naming reveal → "China now has the world's number one model too; its name is Seedance 2.0."

Formula 9 scene opening and Formula 11 interrogative sentence do not belong to the main force of any task, they are "scarce weapons"; see step three.

Step 3: Produce candidates (give 8-12 at a time)

Don't give only one. Output according to the following ratio:

3 main task group(use the main task selected in step one + the corresponding main formula)
3 secondary task group（rewrite with a different main task—informing switches to witnessing, promising switches to judgment. Switching tasks for the same material makes the skeleton completely different; this is the most effective way to find differences)
2 scarce weapons(Formula 9 scene opening / Formula 11 interrogative sentence; these two types are used only a few times a year by Kazike)
2 conservative versions(information given directly, no fancy tricks, used when the user finds the previous ones too aggressive)

After each one, use one sentence to note which formula was used and where the hook is.

Step 4: Self-check

Read references/checklist.md, and go through the first layer of hard indicators item by item (word count/ending/em dash/punctuation).

Must be automatically corrected：

If there is "——" → replace with a comma
If there is ":" → replace with a comma
If the ending is not a period → add a period
If word count >35 → cut to within 30 characters
If more than 2 numbers appear in one item → delete the extra ones

Then givetop 3 recommendations, and explain the reasons for recommendation (which one is most likely to be clicked, and why).

Output format

## 标题任务
[主任务判断 + 一句依据：主干动作是"发生了什么 / 能拿到什么 / 我做了 / 我认为"]
[副任务：叠了哪一类骨架做钩子]

## 标题候选

### 主任务组
1. 「标题」  → 公式X + 公式Y | 钩子：…
2. …
3. …

### 换任务再写一遍
4. …  → 主任务改为「见证」，公式4 + 公式3
5. …
6. …

### 稀缺武器（少用才有效）
7. …  → 场景开场式，适合当"这两天有个事"的系列开场
8. …  → 疑问式，卡兹克一年只用 10 次

### 保守版
9. …  → 信息直给，不玩花活
10. …

## 推荐
第一选：「…」——理由
第二选：「…」——理由

## 自检结果
- 字数：全部 20-35 字，中位 XX 字 ✅
- 标点：0 破折号 / 0 感叹号 / X 问号 ✅
- 结构：X 条叠用了两种骨架
- 避坑提示：[本月已用过"封神"X 次之类的提醒]

Migrating to non-Kazike accounts

This set of formulas is the universal "Chinese internet viral title structure," but if the user operates their own account (such as a technical tutorial account, workplace account, or maternal and infant account), two adjustments need to be made:

Keep the structure, replace the vocabulary. Formula 3 contrastive twist, Formula 5 phenomenon assertion, and Formula 6 numbered list are universal structures and do not need to be changed. Formula 4 first person only requires that "you really did it yourself," with no field restriction.
Replace brand nouns with those from your own field。Kazike's 50.5% of titles contain AI product names, because he is an AI blogger. If it were a finance account, this position would be replaced with company names/fund names/policy names.
Don't copy his catchphrases。悄悄、居然、邪修、封神 is Kazike's personal marker. Using it too much turns into an imitation show. Just pick 2-3 to use.

Reference materials

references/formulas.md — full breakdown of the 12 formulas (skeleton slots / real samples / why they work / pitfalls to avoid)
references/checklist.md — four-layer self-check list + high-frequency failure points
references/title_library.md — complete title library (307 entries, grouped by topic, with skeleton tags).This is a locally generated data file and is not distributed with this repository— out of respect for the original author's copyright, the repository only keeps a small number of examples for each formula. If you want the complete title library to compare and get a feel for it yourself, run scripts/ The script under it can regenerate it.