AI Xiaohongshu Viral Post SOP

You've probably seen the tutorial: "Use AI to batch 100 Xiaohongshu posts a day, make 10k a month." So you paid for ChatGPT Plus, downloaded Jimeng and Kling, ground out 20 image-text posts over a weekend, hit publish, then stared at the backend waiting for the blowup. Three days later, 9 of 10 posts have single-digit views, and the one that broke 100 was a friend's accidental like.

Don't doubt your life-you just hit Xiaohongshu's cruelest wall: its recommendation algorithm identifies and suppresses "AI-feel" content far more aggressively than WeChat. WeChat readers swipe two screens before leaving; Xiaohongshu users scroll past in 0.5 seconds in the feed.

Not my speculation. Over the past half-year I've watched a wave of "pure AI repost accounts" rise and die, usually within two weeks. But I've also watched another group, using the same AI, steadily push image-text posts to thousands of likes. The difference isn't the tool, it's the craft-they use AI as the "assembly-line worker" and themselves as the "QC inspector" and "soul injector." The SOP below splits those two jobs into 5 repeatable steps.

One thing to settle first: Xiaohongshu and WeChat have different underlying logic. WeChat is "media logic"-readers follow accounts and columns; Xiaohongshu is "persona logic"-readers follow "a living person." So using AI on Xiaohongshu carries one more hurdle than on WeChat-you must not only give the content information density, but also make readers feel "this is a person I'd follow." That's why this SOP welds "real experience" into every single step, letting none off the hook.

1. Topic Selection: Don't Brainstorm-Deconstruct Benchmark Hits

The biggest newcomer mistake is sitting in front of the screen inventing topics. 99% of your "inspiration" is angles others wrote to death. The right move on Xiaohongshu is "benchmark deconstruction": find 3-5 hits from your track in the last 30 days, break them into a skeleton, then migrate that skeleton onto your real experience.

How to find benchmarks? Don't doomscroll the Explore feed. Search your track's core keyword on Xiaohongshu, switch to "Hottest" sort, filter time to "past month." Pick ones with a healthy like-to-save ratio-where saves approach or exceed likes, the content has "retention value," a practical image-text hit; high likes but low saves means it's emotional, hard to replicate.

After picking 3, don't rush to imitate-have AI do a structured deconstruction first:

You are a Xiaohongshu operator with 3 years of experience, skilled at breaking viral posts into reusable topic skeletons.
I'll give you the body text of 3 benchmark viral posts. Deconstruct them per the structure below.

【Benchmark posts】
Post 1: {{post 1 body}}
Post 2: {{post 2 body}}
Post 3: {{post 3 body}}

【Deconstruction dimensions, output as a table】
1. Target audience and core pain point (one sentence)
2. Hook type (counter-intuitive / pain scare / result promise / curiosity gap-pick one and explain the trigger)
3. Title formula (extract into a template, mark replaceable slots with {{variables}})
4. Body structure (how many sections, each section's function, emoji rhythm)
5. High-frequency questions in the comments (at least 5, representing real needs)
6. Whether this angle can migrate to my track {{my track}}, give 3 actionable migrated topics

The last column must give "the concrete way to write it migrated to {{my track}}," not vague generalities.

Don't take the deconstruction at face value. I run every migrated topic through a "three-question" filter: first, have I personally done this and hit pits? If not, cut it. Second, does the hook still hold when rewritten in my voice? If not, it depends on the original author's specific persona and copying it will expose you. Third, can I shoot the images myself or stably generate them with AI? Xiaohongshu is image-led-text; if the images aren't doable, drop the topic no matter how good. After the three questions, a table full of candidates usually shrinks to 5-6 you can actually execute.

Once deconstructed, you'll find viral angles are really just a few types: counter-intuitive, pain resonance, result promise, curiosity gap. You don't need to invent new angles-you need to fill a proven skeleton with your real experience.

2. Titles: The "Info Gap" Engineering Within 20 Characters

Xiaohongshu titles are crueler than WeChat's-you get 20 characters, and you're competing against dozens of thumbnails in the feed. People fail in two extremes: either the title is a summary ("My diet meal share"), flat as water; or it's piled with stale memes like "juejuezi," "yyds," "really thank you," which the algorithm flags as low-quality marketing.

A good title's essence is creating an "info gap" that makes readers uncomfortable not clicking. I run this prompt every time, having AI produce 3 titles per formula across 6 formulas-18 total-then manually filter:

You are a Xiaohongshu title writer. Output only titles, no explanation, no fluff.
My track: {{track}}
The core pain point this post solves: {{reader pain point}}
The real result or experience I can offer: {{my real result, e.g., "3 months from 0 to 5k/month"}}

Generate 3 titles for each of the 6 formulas below, 18 total:
1. Number + pain: "The {{N}} {{methods}} 99% of {{audience}} don't know"
2. Counter-intuitive: "Stop {{common practice}}-this is the right posture for {{result}}"
3. Identity resonance: "As a {{identity}}, I was wrecked by {{pain point}} for {{time}}, until..."
4. Result promise: "From {{start}} to {{result}} in {{time}}, I only got these {{N}} things right"
5. Curiosity gap: "The truth about {{result}} no one wants to tell you"
6. Contrast: "Same {{action}}, why did she get {{result A}} and I got {{result B}}"

Requirements:
- Each title ≤ 20 characters (Chinese count), may add 1-2 emoji (at the end, don't pile them)
- Banned: "juejuezi," "yyds," "really thank you," "who gets it, family"
- Numbers must be real and deliverable, not exaggerated
- Leave at least one "unfinished info" in the title that forces a click

My dumb filtering trick: drop the 18 into a 5-person group chat, have everyone blindly pick "the 3 you'd most click." Friends' votes beat my gut 10x. Then I manually swap a word or two, replacing "me" with a more specific identity ("Shanghai-drift worker" beats "me" for clicks).

A counter-intuitive reminder: don't let AI pick "the best one" for you. AI's taste in titles skews toward the safe and comprehensive-exactly the kind Xiaohongshu rewards least. The one it calls "meh" is often the one with the biggest info gap and the one you should use. The final call must stay with a human, ideally a friend who doesn't know your track-whether they click is the real user reaction.

3. Body: A Fixed Five-Section Structure + Functional Emoji

A Xiaohongshu body is not a long WeChat article. Readers have no patience; the first 3 sentences decide life or death. I use a fixed five-section structure, have AI fill the frame, then stuff in real details myself:

You are a Xiaohongshu viral body writer. Write a {{word count, e.g. 600}}-character image-text body in the fixed structure below.

【Persona】{{my identity + a real experience, 2-3 sentences, with emotion}}
【Core pain point】{{the one problem the reader most wants solved}}
【Practical points】{{3-5 real methods, each with my execution details and small data}}

【Body structure, must follow this order strictly】
1. Hook opening (1-2 sentences): create an info gap or contrast. Banned: "today I'll share with you," "hi sisters"
2. Pain resonance (2-3 sentences): articulate the reader's pain so they feel "this person is talking about me"
3. Practical points (3-5): each uses "conclusion first + my execution + one counter-intuitive detail"
4. Pitfall reminder (1 paragraph): 1-2 pits I personally hit, with the real cost
5. Interactive ending (1 sentence): pose a choice or ask for help, guiding comments

【Emoji rhythm】
- 1 relevant emoji at the start of each paragraph, don't pile them mid-paragraph
- No more than 3 emoji in a row
- Use functional: 📌⚠️💡👉✅, not decorative: 🎀✨💕🥺
- Emoji doesn't replace punctuation; commas and periods stay

【Banned words, rewrite if they appear】
"in today's," "it's worth mentioning," "to summarize," "not only... but also," "as is well known"

【Sentence style】
- Each sentence ≤ 25 characters, favor short sentences and line breaks
- Each paragraph ≤ 3 lines, so 2 paragraphs fit on one phone screen

The key here isn't making AI write beautifully, it's keeping it from drifting. After generation I do two things: first, rewrite every "golden line" AI produced in my own voice, adding an "I" and a real detail; second, cut 30% of the words-AI always overwrites by a third, and Xiaohongshu's density demand is "every screen has new information."

A real rewrite example: AI gave me "reasonable time management can significantly improve personal efficiency"-standard drivel. I changed it to "I tried 5 time-management methods; only the Pomodoro Technique survived two weeks, the rest all died on day 4." Same meaning, but the second has a person, a number, and a failure-AI can't fabricate that. I now do 8-10 such rewrites per post; it's the most labor-intensive and most valuable step in the whole flow.

4. Images: The AI Image-Gen Toolchain and Cover Engineering

Xiaohongshu is an "image-led-text" platform; the cover determines 70% of click-through. Many generate a pretty AI image and post it, only to find the text is blurry, the ratio is wrong, and half of it is cropped in the feed. For images, AI handles "image generation," you handle "engineering."

Cover trinity: 3:4 portrait (1080×1440px), main title occupying 1/3 of the frame, high-contrast color. Body images run 3-6, the first is the cover, the last is an "interaction guide image" (writing "Which step are you stuck on? Tell me in the comments").

You are a Xiaohongshu cover designer + AI image-gen prompt engineer.
I want a set of images for "{{post topic}}": a cover + 3 body images.

【Cover requirements】
- Size 3:4 (1080×1440px)
- Main title "{{main title}}" occupies 1/3 of the frame, bold sans-serif, white or yellow background for high contrast
- Subtitle "{{subtitle}}" one small line, directly below the main title
- Style: {{clean minimal / magazine / journal / before-after comparison}}
- No more than 2 lines of text on the image, don't pile text

【Output】
1. Cover layout description: text coordinate positions, font-size ratio (main vs sub), hex colors
2. AI image-gen prompts (both Chinese and English versions), including: subject, lens, lighting, composition, style keywords, negative prompts
3. Image-gen prompts for 3 body images, corresponding to the 3 practical points of the body

【Tool priority】
Jimeng / Kling (Chinese-friendly, more stable text rendering) > Midjourney v6 (better texture but Chinese text garbles) > SDXL (controllable but needs tuning)
Mark which tool each prompt is recommended for, and note that tool's text-rendering pitfall.

A counter-intuitive point: the biggest pitfall of AI image-gen isn't ugly images, it's "the text on the image." Jimeng and Kling now render Chinese text more stably than Midjourney, but still produce wrong characters. My approach: have AI generate only the background and subject, then add all text in via Chuangketie, Gaoding Design, or Photoshop afterward, ensuring clarity and controllable layout.

I generate 3 cover versions at once, run them through a blind vote in an alt account or friends' group, then finalize. Don't迷信 "first-glance instinct"-the cover I think looks best often loses on click-through to the one I thought was "too plain." Data is more honest than taste. If the last post's numbers were mediocre, next time I'll swap only the cover and keep the body untouched, as a controlled experiment-that isolates the problem far better than changing multiple variables at once.

5. Publish Optimization: Hashtags, Timing, Comment Seeding

Done writing, just hit publish? That's the most costly mistake for newcomers. Xiaohongshu's cold start relies on interaction data in the first 2 hours after publishing (click-through, read-completion, comments, saves)-you must seed before publishing.

You are a Xiaohongshu publish-optimization assistant. Based on the post below, help me complete three pre-publish tasks.

【Post body】
{{full body text}}
【My track】{{track}}
【Target reader profile】{{e.g., 25-30 working women in tier-1/2 cities}}

Output:
1. Hashtags (5-8): ratio of "1 big-traffic word + 2 precise mid words + 2-3 long-tail conversion words," each prefixed with #. Don't piggyback on trending tags unrelated to the body-it gets demoted for content mismatch
2. Publish-time suggestion: give 3 optimal slots this week, each explaining why (tied to the target reader's active scenarios: commute, lunch break, before bed), with timezone noted
3. Comment seeding (3 seed comments): ones I post via an alt account or have friends post, each either supplementing a practical detail or posing a follow-up question to spark discussion-not bot-flavored "so great"
4. High-frequency comment reply templates (2): for replies to "need the method" / "saved" type comments, designed to get the other person to share more or follow me for the follow-up

A timing note from experience: don't迷信 "8 PM has the most traffic"-that's when every account piles in, and newcomers get drowned by big accounts. Finding your target reader's niche active slot is more effective, e.g. career content at 7:30 AM on the subway commute, parenting content at 10 PM after the kid's asleep.

Another cold-start detail that gets ignored: in the first 30 minutes after publishing, reply to every comment as fast as you can. The algorithm watches interaction "liveliness" in real time; reply fast and it judges the post has discussion heat, then adds exposure. I set a hard rule-stay on the phone for 1 hour after publishing, reply to comments within 5 minutes, and always tack on a question to keep the conversation going instead of just "thanks."

6. Pitfall List: 9 Pits Every Newcomer Hits

1. Piling title emoji with stale memes. "Juejuezi," "yyds," "really thank you" are now flagged by the algorithm as low-quality marketing signals-the more you pile, the harder the throttle. One functional emoji beats five decorative ones.

2. Using reposted web images or unlabeled pure-AI images. Xiaohongshu polices originality harder than WeChat; repost detection directly throttles or even bans. AI images need secondary processing (add layout, text, collage), and you must truthfully mark "AI generated" in the creation statement.

3. Not re-cropping cover text. AI-generated images often have blurry or feed-cropped text. Always crop to 3:4 and separately verify the title is fully visible-don't let half a word show in the feed.

4. Pasting WeChat-style long articles into the body. Xiaohongshu readers have a third of WeChat's patience; long paragraphs drive them off. Split long sentences short, break paragraphs up, at least 2 paragraphs per screen.

5. Frequent delete-and-republish after posting. The deadliest one-post gets no data, you delete and repost, the system flags abnormal behavior and drops your weight. Don't touch it for 48 hours; if there's no traffic, review the cover and title, then iterate on the next post.

6. Stuffing WeChat IDs for traffic on day one. A new account hasn't built trust; jamming a WeChat ID into the body and comments triggers an instant marketing-violation throttle. Build account weight and followers first, then use the bio or the "fan group" feature for compliant handoff.

7. Running a multi-account matrix without device isolation. Logging multiple accounts on one phone, posting similar content on the same IP, gets flagged as linked marketing-one falls, all fall. Matrix accounts must use separate devices and IPs, or just focus on doing one account well first.

8. Only watching the recommendation feed, ignoring search long-tail traffic. Xiaohongshu's search traffic lasts far longer than WeChat's; a practical post can still get found half a year later. The catch: your title and body must embed keywords readers actually search. Don't fixate on the 48-hour recommendation burst-burying one high-frequency search term in the title can stretch a post's lifespan from days to months.

9. Chasing trends, zigzagging across tracks. Career today, beauty tomorrow, relationships the day after-the algorithm can't tag you, and recommendation traffic never starts. Stick to one niche track for your first 30 posts; letting the system know who you are and who to push you to beats chasing any trend.

References

• Xiaohongshu Creator Center (post data, follower profile, creation statement): https://creator.xiaohongshu.com/
• Xiaohongshu official site (community guidelines and post-publish rules): https://www.xiaohongshu.com/
• Xiaohongshu Explore feed (benchmark topics and trend reference): https://www.xiaohongshu.com/explore
• Pugongying Brand Collaboration Platform (compliant deal-taking and filing): https://蒲公英.xiaohongshu.com/

Every prompt in this piece, I still use today. AI didn't turn me into a viral-post machine, but it moved me from "post by feel" to "post by process"-and on Xiaohongshu, avoiding pits is the biggest dividend there is.