中/EN
Field SOP
Field SOP

Content Repurposing Workflow: One Source into X/Xiaohongshu/WeChat/Zhihu Variants

Feed one source article into n8n; the LLM rewrites it per X/Xiaohongshu/WeChat/Zhihu specs and archives all four to Feishu/Notion for human sign-off. Includes n8n JSON template, four platform-specific prompts, and six pitfalls.

Published July 29, 20266 min read
<!-- workflow-content-repurposing | resource | Content Repurposing Workflow: One Source into X/Xiaohongshu/WeChat/Zhihu Variants -->

You finish a long technical piece, then manually re-chop it into X, Xiaohongshu, WeChat Official Account, and Zhihu versions-each with its own length, tone, and tag rules. Copy-paste-format-tweak eats a whole evening. X caps at 280 characters, forcing you to compress a 3000-word piece into one opinion; Xiaohongshu wants emoji bullets, a hooky opener, an interactive closing question; WeChat wants H2 sections with cases and steps; Zhihu wants conclusion-first with counter-arguments and real citations. Switching all four rule sets in your head, by version three you start phoning it in.

Worse, mid-way you slip the same paragraph into both Xiaohongshu and Zhihu, and both flops-Xiaohongshu calls it too dry, Zhihu calls it too thin. There's a hidden problem too: rewriting the same source four times, your memory of the content drifts by the later versions-Xiaohongshu might drop a key data point from paragraph two, WeChat might turn "improved 30%" into "significantly improved", and you only catch it after publishing.

This n8n workflow takes one source article, runs it through LLM rewrites that follow each platform's real spec, and drops all four variants into Feishu Bitable or Notion for human sign-off. The core value isn't saving time-it's completing all four rewrites in one pass, avoiding the memory drift that creeps in when a human rewrites the same thing four times; the author only handles raw source and final approval, the format-shuffling goes to the machine. Solo content creators and ops team members handling multi-platform distribution benefit most-the more rewrites, the more time saved.

Workflow Chain

Manual trigger -> Set node pastes source (title + body) -> four parallel LLM rewrites (X / Xiaohongshu / WeChat / Zhihu, each its own prompt) -> four parallel writes to storage (Feishu Bitable rows or Notion pages) -> human review before publishing.

Download Template

Content Repurposing Workflown8n JSON

Setup Steps

  1. Import: n8n -> Workflows -> Import from File, pick workflow-content-repurposing.json
  2. Paste source node: fill source_title and source_text in the Set node. Read the source once first and cut irrelevant sections-the LLM won't decide what's core for you, it rewrites whatever you feed it; cutting clutter is human work
  3. Four LLM nodes: one prompt per platform, already in the template. Fill your API key; DeepSeek/Kimi/Qwen all work. For long sources (3000+ words), switch to a 32k context model-8k truncates the tail
  4. Model choice: rewriting is instruction-heavy, don't use the smallest cup. DeepSeek-V3, Kimi K2, Qwen-Max tier is fine; run all four branches on the same model to avoid tone drift
  5. Archive node: Feishu Bitable via HTTP Request + OpenAPI, four rows with platform + content fields; Notion via the official node, one child page per variant under a "repurpose pool" database. Add a status column defaulting to pending, flip to approved after review so you can batch-filter unreviewed ones
  6. Test run: trigger manually, feed in a recent long piece of yours, check whether the four outputs match each platform's tone. First run usually needs 2-3 prompt tuning rounds-focus on whether the X variant hits 280 chars and whether Xiaohongshu's emoji and tags land
  7. Human sign-off: don't ship LLM output directly. X-check facts aren't compressed into distortion, Xiaohongshu-check medical/finance sensitive words, WeChat-verify data sources, Zhihu-check citations are real. At review, put the source on the left, the rewrite on the right, and diff paragraph by paragraph-focusing on numbers and causal claims

Companion Prompts (One per Platform)

X variant (280-char hard limit)

Prompt
You are an X(Twitter) rewriter. Compress the source to under 280 characters.
Rules:
- First 80 characters build a hook, open with contrast or conclusion
- Keep 1 core fact, drop the rest
- End with 1 relevant hashtag
- Tone: direct, opinionated, debatable
- No emoji piling, 1 max
Output plain text only, no explanation.
Source title: {{source_title}}
Source body: {{source_text}}

Xiaohongshu variant (under 1000 chars)

Prompt
You are a Xiaohongshu rewriter. Output under 1000 characters.
Structure:
1. Opening hook with pain point or contrast, under 20 characters
2. 3-5 points, each with one emoji
3. First-person voice, "I tried it", "been there"
4. End with an interactive question to drive comments
5. 3-5 # hashtags
Tone: friend recommending, lived-in.
Compliance: no medical absolutes ("cure", "best"), no financial returns promises ("guaranteed", "risk-free").
Output plain text only.
Source title: {{source_title}}
Source body: {{source_text}}

WeChat Official Account variant (1500-2500 chars)

Prompt
You are a WeChat Official Account rewriter. Output 1500-2500 characters.
Structure:
1. Hook with scene or data, under 200 characters
2. Body split into 3-4 H2 sections
3. Each section pairs with steps or a case
4. End with summary + follow prompt
Tone: professional but not stiff, with author voice.
Keep core facts from the source, do not fabricate data or citations.
Output Markdown only.
Source title: {{source_title}}
Source body: {{source_text}}

Zhihu variant (~2000 chars, Q&A style)

Prompt
You are a Zhihu rewriter. Output around 2000 characters, Q&A style.
Structure:
1. Open with a one-sentence conclusion
2. 3-4 layers of argument, each with data or source
3. Add a counter-perspective ("but note that")
4. End with actionable advice
Tone: rational, information-dense, like a serious answer.
Citations must be real-no fabricated papers, names, or links.
Output Markdown only.
Source title: {{source_title}}
Source body: {{source_text}}

Pitfalls

  • Don't one-size-fits-all the tone: using the same prompt across all four and only changing word count is the top failure mode. Xiaohongshu wants "I" first-person plus emoji, Zhihu wants conclusion-first plus counter-argument, X wants sharp opinions-tone differences aren't covered by length, you need four independent prompts
  • Hard length limits truncate facts: X's 280 chars and Xiaohongshu's 1000 are real hard limits-exceed and you can't post or it folds. But the LLM will cut key facts to fit. Telling it "keep 1 core fact" is more controllable than "summarize the whole thing"-you want a hook, not an abstract
  • More compliance words than you think: Xiaohongshu auto-blocks medical ("cure", "eradicate"), financial ("risk-free", "guaranteed return"), and absolutes ("first", "best"); Zhihu also limits unlabeled ads. Listing banned words in the prompt beats publishing, getting throttled, then editing. Also, Xiaohongshu's sensitive-word list updates-rescan outputs quarterly and add frequently-blocked words back into the prompt
  • Human sign-off is not optional: the most common LLM rewrite bug is "fact compression distortion"-turning "improved 30%" into "significantly improved", "a study" into "experts say". Especially on X and Xiaohongshu short versions, every word is a trade-off; at review, diff against the source line by line, focusing on numbers, percentages, and causal claims-those three lose precision fastest when "summarized"
  • Same model across all four branches prevents drift: four branches on four different models fight each other-WeChat on DeepSeek stays steady, Xiaohongshu on Kimi suddenly gets literary. For one repurposing batch, run all four on the same model; switch the whole batch if you must switch. Another reason: four parallel branches have different latencies, and uneven speeds scramble archive ordering-same model at least keeps the variance small
  • Long sources need chunking: if the source exceeds 4000 words, an 8k context model can't take it all. Don't say "ignore the overflow" in the prompt-the LLM silently drops the tail, and you don't know which part. Two fixes: either switch to a 32k model, or add a Code node before the Set node to split by paragraph, rewrite in batches, then stitch back

The template is the skeleton; how you paste source and whether you archive to Feishu or Notion is yours. After it runs, add two nodes: first, run the source through an LLM "fact summary" before rewriting, then after all four variants land, scan them back against the summary to catch lost facts; second, auto-create a "pending review" task in Feishu todo after archiving, reminding you to sign off same-day so rewrites don't pile up and gather dust.

References

This article is AI-assisted and human-edited. Last updated: 2026-07-29

Related