One Topic Can't Be Copied Into a Xiaohongshu and a Douyin Version

Posting the same topic twice isn't cross-platform

My old approach was straightforward: write one full article, then have the model shorten it into a Xiaohongshu version and a Douyin version. Fast — but after posting, you could barely tell the two platforms apart.

Xiaohongshu looked like a WeChat article cut into pages; Douyin looked like someone reading the article aloud. The cover, title and ending repeated the same line, and when the numbers were bad, I couldn't even tell whether the topic was off or the content simply hadn't been redesigned for the platform.

Later I started saving what could be shared up front, and pushing the platform deliverables to the very end. Facts, judgments, assets and steps can be shared; what the user sees in the first second, the order information appears in, and what they do afterward must be decided fresh.

That's the dividing line between two pieces — not word count.

The intermediate artifact can't be a "generic article"

If the middle layer is already a full article, later conversion can only trim around the original structure. To leave room for platform-native redesign, I use a fact pack, not a master draft.

A fact pack includes at least:

• The problem: where the reader is stuck.
• Verifiable facts: source, date and applicable conditions.
• A demo action: what can actually be shown on screen.
• Failure notes: which step might go wrong.
• The main judgment: what Kenton ultimately recommends.
• An asset list: screenshots, screen recordings, diagrams and their copyright status.
• Forbidden claims: effects without evidence can't be written.

It doesn't contain platform language like "hey everyone, today I'll teach you" or "save this," and it doesn't pre-decide how many images or how long the video.

Seeing two structures through one topic

Say the topic is "let AI tidy up a messy folder." The fact pack has three real actions:

1. List the files first, don't move anything.
2. Have AI propose a categorization.
3. After a person confirms, then rename or move.

When the same facts enter Xiaohongshu and Douyin, the deliverable shouldn't just change length.

The Xiaohongshu version: the user is scanning images

The first image should convey the problem and the result, e.g. "Have AI list the files first — don't start by moving them."

Later images can unfold in reading order:

• Image 2: why moving directly is risky.
• Image 3: which fields the input list needs.
• Image 4: the prompt that makes AI only propose a plan.
• Image 5: which conflicts a person confirms.
• Image 6: how to tell the tidy-up is done.
• Last image: save a reusable checklist.

Each image carries one independently understandable node of information. The images aren't a backdrop for the body — a user who only sees the first two should still know whether this piece suits them.

The Douyin version: the user is watching actions happen in sequence

The short-video script has to start from the screen:

On screen: open a directory full of files
Voiceover: don't let AI move anything yet — step one is just export the file list

On screen: the terminal generates the list
Voiceover: this keeps the filenames, types and sizes, and doesn't touch the originals

On screen: AI returns a categorization suggestion
Voiceover: for now it can only propose — before anything moves, a person confirms duplicates and important files

On screen: confirm, then run
Voiceover: once you see the new directory, re-check the file count and the original backup

The voiceover describes real on-screen actions; it doesn't suddenly jump to "reshape your productivity." With the sound off, the screen still makes sense; without watching the screen, the voiceover doesn't invent actions that don't exist.

The title isn't a synonym rewrite

A Xiaohongshu title solves "why should I stop and read this image set" — it can stress the audience, the risk, or a savable result.

A Douyin title solves "what will I see in the next few seconds" — it fits pointing at an action or a conflict.

The same topic could become:

Xiaohongshu: When your files are a mess, have AI make a list first — don't tidy directly

Douyin: I let AI organize my folder, but my first command forbade it from moving files

The two titles share the same facts but enter different use scenes. Neither can promise "automatically sort ten thousand files in one minute" unless the content actually does and proves it.

My two-platform production order

To stop the platforms from contaminating each other, I produce in this order:

1. Save the source and the real operation record.
2. Extract the fact pack, with no platform tone.
3. Settle on a single main judgment.
4. Choose the image-and-text nodes Xiaohongshu needs.
5. Choose the actions Douyin can actually film.
6. Write the title, first screen and ending separately.
7. Do platform risk checks independently.
8. Save the publish package and platform read-back separately.

Automation is best at the data-organizing of the first two steps, and the format checks at the end. The structural choices of the platform deliverable still need to be made against the real assets.

The two platforms' acceptance questions also differ

For the Xiaohongshu version I ask:

• Whether the topic is still readable when the cover shrinks.
• Whether the first two images lay out the problem and the core judgment.
• Whether each image carries only one main task.
• Whether the body supplements conditions and risks the images couldn't hold.
• Whether a user who saves it can reproduce the steps in order.

For the Douyin version I ask:

• Whether the first second already has real on-screen information.
• Whether each line of voiceover matches the current action.
• Whether a key precondition was cut for the sake of pacing.
• Whether, on mute, the subtitles and screen work independently.
• Whether the ending action relates naturally to the content.

Checking both platforms with the same "generic QA sheet" misses the most important problems.

Automation failure usually happens in the middle layer

If the two versions still look like copies, first check whether the fact pack was secretly written as a full article. Common signals:

• It already has a fixed opening and ending.
• All the information is locked into one narrative order.
• The asset list only says "add some nice-looking images."
• No record of the specific actions that can be filmed.
• The conclusion appears before the evidence.

The fix isn't adding another "make it more platform-native" — it's going back to the fact pack and separating facts, judgment, steps and assets.

Don't use the number of platforms to prove automation ability

One input generating five pieces of copy looks efficient, but if all five have to be redone, automation just manufactured the rework ahead of time.

I'd rather stably nail one platform, then distill the shareable fact layer. When the second platform has real assets and a clear entrance, add its own Writer and acceptance rules.

What's reusable across platforms is the evidence, the facts and the assets — not the finished sentences. Split your most recent piece into two files, a "fact pack" and the "platform deliverables," then see whether Xiaohongshu's first image and Douyin's first action are still the same. If they are, the problem is in the middle layer.