xiaohongshu-article-packager
Xiaohongshu Article Packager
Follow shared release-shell rules in:
postplus-sharedrelease-shell rules
Use this skill when the user wants:
- 一篇原稿变成可发的小红书图文
- 口播稿改成图文分页
- 给一段观点内容做分页、层级、配图和可渲染页面
- 产出适合后续排版或渲染的结构化内容包,但保留原文
This skill is for layout packaging.
It is not for:
- platform scraping
- competitor research
- direct posting
- fabricating claims not present in the source
- rewriting the source by default
If the request is mainly about platform research, use skills/20-research/xiaohongshu-research.
Core Idea
Do not turn the source into a different piece of writing.
The job has two layers:
structure without rewrite- split the source into readable pages
- choose which original lines deserve larger emphasis
- preserve the original sequence unless the user asks to reorder
render packaging- assign page roles
- place images
- prepare clean layout instructions or HTML/CSS for rendering
The default goal is:
- readable
- faithful to the original wording
- faithful to the original information density
- faithful to the original voice
Inputs
Typical source inputs:
- a markdown file
- a long paragraph dump
- a script written for talking-head video
- rough bullets
- a merged note with too much information
Useful optional inputs:
- preferred page count
- image assets to place
- desired visual direction
- whether the output is: carousel pages or one long image
If these are missing, infer the smallest safe set from the source.
Output Contract
Default outputs should be a package with these sections:
Source Rule- state that the wording is preserved
Cover Copy- only use original lines or original line fragments unless the user explicitly allows rewriting
Page Plan- 6 to 10 pages by default
- one clear purpose per page
Page Copy- source text assigned to each page
Layout Brief- structured instructions for downstream layout
Render Files- optional HTML/CSS when requested
When the user asks for files, prefer:
01-outline.md02-titles.md03-pages.md04-caption.md05-layout-brief.md
Read references/output-template.md when composing the package.
Default Workflow
1. Respect the source as immutable by default
Unless the user explicitly asks for rewriting:
- do not paraphrase
- do not compress
- do not add summary lines
- do not add CTA lines
- do not add a separate caption
Allowed operations:
- split into pages
- promote original lines into headline size
- move image modules around the source text
- add neutral page labels
2. Decide the page architecture
Default to 6-10 pages.
Each page should do one job only:
- cover
- tension
- key distinction
- proof or explanation
- consequence
- closing or CTA
Do not use page breaks mechanically every N characters. Split on natural paragraph or sentence boundaries from the source.
3. Write for page readability without rewriting
For each page:
- keep one main point
- give it a visible title or emphasis line only by promoting original source lines
- avoid dense wall-of-text blocks
- make the first two lines legible without context from the previous page
If a page exceeds reasonable density, split it.
4. Prepare layout-aware copy
Assume the downstream renderer or layout model benefits from structure.
For each page, clearly separate:
pageRoleheadlinefrom source if usedsubheadif neededbodyfrom sourceemphasisLinesfrom source if used
Then prepare a consolidated layout brief.
Read references/layout-system.md when the user wants the warm-gradient card style from the referenced PDF.
Quality Rules
1. Source-first
Do not invent authority, case studies, metrics, or personal stories not present in the source.
2. No stealth rewriting
Do not silently change:
- sentence rhythm
- information density
- spoken-language phrasing
- the user's own wording habits
3. Avoid AI-shaped cleanup
Do not force the source into:
- neat binary formulas
- over-organized summary language
- your own preferred phrasing patterns
If the source says it one way, keep that way unless explicitly asked to rewrite.
4. Cover copy must be from the source
The cover should use:
- original lines
- or original line fragments
not a fresh headline invented by the model.
Heuristics
Good cover material
- strong original opening lines
- strong original judgments
- strong original rhetorical questions
Weak cover material
- lines that only work after a lot of setup
- lines that require new explanation from the model
Common Input Shape
A common use case is:
- a draft in the active work folder or a nearby project directory
- especially talking-head scripts or founder/operator viewpoint drafts
- output needs to feel publishable on Xiaohongshu without losing the original opinion
For example:
口播稿 -> 小红书图文分页观点长文 -> 封面 + 标题 + 8 页图文创始人表达 -> 更适合图文阅读的版本
Default File Placement
If the user asks you to write outputs to disk, prefer creating a sibling folder next to the source draft, for example:
<source-dir>/<draft-name>-xhs-package/
Store the package files there using the default filenames from the output contract.
Release-Shell Execution Contract
- keep page planning notes, source extracts, and intermediate packaging files
under
<work-folder>/.postplus/xiaohongshu-article-packager/ - keep only final user-facing page packages outside
.postplus/ - start with a bounded first pass on one draft and one package target before broader packaging
- if the source is insufficient or under-specified, stop immediately instead of inventing missing material or relying on customer-specific path assumptions
Failure Modes
Stop and say the source is insufficient if:
- there is no clear argument
- the draft is only slogans
- the user expects claims or case studies not in the material
Then offer one of:
- a lighter layout pass
- a request for more source material
- a narrower page selection based on what is actually present
More from postplusai/postplus-skills
audio-transcription
Transcribe local or remote audio into durable text and timestamp artifacts using hosted Whisper models. Use this when the job is speech-to-text from audio files and you need request/response persistence, optional timestamps, and subtitle-ready outputs.
82google-trends-research
Research Google Trends search-intent signals for topic discovery, keyword momentum, regional interest, and rising queries without treating search trends as the same thing as platform content heat or marketplace demand.
77seedance-submitter
Use when preparing, submitting, polling, or debugging Seedance 2.0 video generation jobs from product images, storyboard images, UGC scripts, voiceover copy, or promptPlan request JSON. Use for splitting scripts into render segments, uploading references, creating request JSON, submitting jobs through the hosted capability, polling predictions, and handing off local render paths.
75social-media-publisher
Prepare and, after explicit approval, publish social posts through the PostPlus platform-owned Postiz workspace.
75facebook-research
Research Facebook pages, public follower or following surfaces, and public posts using hosted collection capability. Use this when the user wants Facebook account research, follower-surface sampling, or public post metrics.
75x-tools
Local execution tools for X/Twitter hosted collection workflows, including actor runs, dataset normalization, tweet ranking, account ranking, audience graph construction, and language clustering.
74