xiaohongshu-skill

Xiaohongshu Image-and-Text Skill

Used to create a Xiaohongshu/Rednote image-and-text note that users can directly accept. By default, confirm the strategy and visual direction first, then proceed to image generation actions; do not treat "image prompts" as image deliverables, and do not read all references right away. The complete 15-section workbench is an archival and verification contract, not the default final chat display.

Core Rules

First look at real materials, audience, boundaries, and user intent; do not start from viral templates. Then research on your own what readers will search for, what scenarios they get stuck in, common angles for similar content, and verifiable high-signal samples. Do not throw "where does the traffic come from" back to the user to answer.
The original user input is the highest-priority task fact; HookPrompt, the optimized full prompt, model-written scope, or summaries must not add restrictions the user never stated, such as "only copy," "no image production needed," or "no images." If the original input conflicts with the optimized prompt, the original input prevails.
This skill has no pure-copy shortcut: once this skill is triggered, it must go through the strategy, title/cover text, visual route, cover MVP, post-confirmation batch generation, and pre-publish self-check process.
The strategy route, title/cover text, visual route, image Brief, cover candidates, and batch image generation permission are all user decision points; ambiguous input must first be completed through Codex request_user_input、Claude Code AskUserQuestion or a host-equivalent native dialog.
Once the MVP of the cover image and cover text is confirmed by the user, it is regarded as production permission for the complete image-and-text package: you must continue to generate at least 8 titles, 1 body copy, a cover, and 6-8 inner-page images; the scope must not be downgraded again to "only write copy/only make a script/only make a cover."
The image route priority cannot skip levels: Image2 / host-native image generation > other image generation MCPs > existing assets > SVG/HTML/local layout. When either Image2 or another image MCP is available, SVG/HTML/local layout is prohibited; only when both are unavailable or both fail is downgrading allowed, and the downgrade evidence and reason must be stated. For the authoritative definition, see references/skill-contract.md"Image Route Contract."
The optional visual workbench is only responsible for carrying the determined 3:4 page slots, inserting the current round's generated assets, thumbnail checks, annotation-revision comparison, and the page-by-page checklist; it is not a new image generation tier, and it cannot impersonate a substitute image generation capability after Image2 fails. For the general contract, see references/visual-workspace-contract.md; specific vendors are mapped only in the adaptation list.
The optional visual workbench is not asked about when the Skill starts. Only after opportunity research, marketing decisions, page-by-page visual direction, and the Image2 Brief are complete, and the first 3:4 cover candidate has actually been generated and the user can immediately compare layouts or annotate, ask once whether to open it; skipping, unavailability, or startup failure all continue the existing cover acceptance and image generation chain.
SVG, HTML cards, and static structural diagrams can only count as structural previews; they cannot impersonate real images and cannot be written as delivery-grade.
Do not write a recommended option as "the user has chosen."human_confirmed It can only come from the user's original words or a host-native selection tool return; otherwise write blocked_pending_human_decision or explicit_auto_permission。
Old local output directories, old images, and historical thread results can only serve as reference or comparison, and cannot count as "generated in this round." Unless the user explicitly requests reusing old images, generation-type requests must have image tool evidence from this round, or honestly write blocked: no_live_image_generation_evidence。
Host-native image generation can serve as an Image2-equivalent primary path; evidence can be imageGeneration、image_gen, newly generated image paths from this round, or host-visible image results. Do not jump to MCP when the host can generate images, and do not write "there is currently no tool named Image2" as a reason to skip native image generation.
For notes related to courses, consulting, services, and knowledge products, avoid traffic-diversion phrases in the public section such as "DM me, join the group, sign up, claim materials, external links, contact information."
The final deliverable must be visible in the chat, but it must be a publish-ready summary: image previews/image list, selected title, alternative titles, body text, continuation boundaries, tags, and missing real materials. Do not dump the full 15-section workbench, strategy process, and prompts into the final chat unless the user explicitly asks for the complete workbench.
When the user has already requested 图文笔记、完整发布包、引流贴/种草贴、封面/内页、图片生成 or "help me generate," the default goal is the complete workflow. Do not set "copy + script, no image generation" as the recommendation again, and do not ask "do you only want the copy" after the upfront decisions are complete.
When the user says "help me generate / make a set / suitable for publishing / image-text / photos or visuals plus text," and does not say "wait for my confirmation first," treat it as first-round cover MVP generation permission: write the status as explicit_auto_permission, and write the source as "the original generation request triggers first-round cover MVP permission." The AI may choose a conservative default strategy, title, visual route, and Brief, and generate 1 cover candidate for acceptance; this permission must not be treated as permission for batch inner pages. Only when the user's exact words clearly say "directly the full set / no need to ask / you decide and complete it fully / directly make the complete image-text package," or cover acceptance returns "approved and generate in batch," is it allowed to generate the cover + 6-8 inner pages.
In new windows, sub-threads, or unattended acceptance scenarios, do not treat "needs user confirmation" as a default reason to stop; only when the host refuses, the image tool fails, the host lacks image generation capability, or the user explicitly requests a pause is blockedallowed. Missing real materials can only trigger an "AI/generated material disclosure" and a conservative Brief, and cannot automatically degrade to only giving image suggestions.
真实女生/美女/真人写真/留学生活/交友 and other realistic expressions are not reasons to stop image generation, nor are they reasons that must be changed to "fictional identity." Create Image2/host-native images according to the realism the user wants; only avoid specific contact information, minors, and illegal or fraudulent elements.
文字信息卡、图片上文字说明是主体、大字封面、知识卡 is not a reason to skip Image2, nor a reason to go directly to local layout. When Image2 or any other image MCP is available, local layout is prohibited; only when there are Chinese typos, unreadability, layout failure, or the user requests editability may it degrade to locally laid-out PNG, and the fallback reason must be stated.
After the user completes upfront choices such as strategy, title/cover text, and visual route, the next step is to execute the selected generation route; do not pop up scope choices like "how far should this go" again. After the user selects "approved and generate in batch" in the cover acceptance dialog, the cover + 6-8 inner page images must continue to be generated, and it is not acceptable to deliver only copy, a script, or a single cover.

Trigger scope

Use this skill when the user wants to create the following:

A complete Xiaohongshu/Rednote image-text note or publishing package.
A complete visual image-text production that explicitly requires a cover and inner pages.
A complete recommendation/content publishing package for courses, services, consulting, products, personal IP, local businesses, and knowledge sharing.

The following do not trigger this skill and are handled as ordinary writing or image capability: a single title, a few sets of cover text, body copy polishing, comment replies, tags, a single drawing prompt, or partial design consultation that does not require a complete publishing package.

First action

Determine whether it is完整发布包. If it is not a complete package, do not trigger this skill, and even more so do not force a partial requirement into the cover MVP process.
First determine from the original user input whether to trigger the Xiaohongshu image-text process; do not downgrade from the inferred scope of HookPrompt. After triggering, first read references/domain-research-brief.md、references/marketing-decision-model.md and references/page-art-direction-system.md。
AI first conducts content opportunity research: readers' search/browsing scenarios, specific pain points, gaps in competing content, deliverable evidence; mark each item as 用户提供、研究证据 or AI 假设. When there is no browsing or retrieval capability, honestly mark the research gap and do not pretend to know popular content.
Then form a marketing decision board and a page-by-page visual director table: each page must clearly state the page task, what to draw, what text to write, where the text goes, why it is placed that way, and the image prompt. Do not first pick a surface style like "real desktop/infographic".
When user intent is ambiguous, use the host's native pop-up to fill in the few pieces of information that would change the direction of the final product: target readers, product/topic, real materials, commercial boundaries. Do not ask the user "where does the traffic come from"; that is a research step.
When strategy, title/cover text, visual route, cover MVP, or whether to generate images would clearly change the result, confirmation must be made using the host's native selection interface; when there is no native selection interface and the user has not explicitly authorized automation, stop at the decision card and do not output a final draft that assumes a choice has been made.
When the user explicitly says "you decide/just do it/no need to ask/you figure it out and execute", you may continue, write the status as explicit_auto_permission, and record the AI's choice rationale and risks; this is not human_confirmed。
The user did not say "just do it", but the original input is already a complete image-text generation action, for example "help me generate a complete Xiaohongshu image-text package" or "make a set of covers and inner pages", and did not ask to stop and wait for confirmation first; in that case you may also enter the first-round cover MVP generation. The status is still written as explicit_auto_permission, and the source must state "the original generation request triggers first-round cover MVP permission", and retain replaceable items, unconfirmed items, and safety statements; this permission only covers 1 cover candidate, not batch generation of 6-8 inner pages.
When the user wants to simulate real materials, directly enter the image Brief and Image2 priority generation process, and do not automatically stop image generation due to lack of real materials.
Before the title/cover text, visual route, and image Brief are confirmed or authorized, do not call Image2; however, a complete generation request itself can serve as first-round cover MVP permission. If entering first-round cover MVP permission, first record the AI default choice rationale, then call Image2/host native image generation to generate 1 cover candidate. After the candidate is visible, first judge according to visual-workspace-contract.md whether it has reached the one-time high-value workbench inquiry node, then use the native pop-up to request "approve and batch generate / modify / stop and review"; the workbench choice cannot replace cover MVP permission.
When images or a complete publishing package are needed, do not use historical outputs/, old generated images, or example files to replace this round's generation. If existing output on the same topic is found, you may only state that "the old package can be referenced", then continue with this round's confirmation and image generation; only when the user explicitly requests reuse of the old package may it be reused.
The definition of the cover MVP is: the cover main title/subtitle, main visual subject, composition text area, mobile thumbnail readability, and batch generation permission are all confirmed at the same time; under the first-round cover MVP permission, AI first sets the MVP default values on its own and produces 1 replaceable cover candidate, finally marked as partial_pending_mvp_confirmation or blocked_pending_human_decision, and must not be marked as human_confirmed or complete。
After completing the batch images, the final chat reply follows the "satisfactory delivery form": a one-sentence process summary, cover preview, title alternatives, body text, follow-up method, tags, image thumbnails/list, and archive files. The complete 15 sections are only written to the file or shown when the user requests it.
The decision popup only asks about choices that will change the creative direction. If the user has already requested a generation-type deliverable, after the upfront decisions are completed, the next step should be "generate cover candidates and prepare the full image set," not a scope downgrade choice of "copy + script / generate cover / copy only."
There is no "copy + script" recommended route. After this skill is triggered, the default recommended goal is to first generate the cover MVP, and after the cover passes, fully generate the cover + 6-8 inner pages; when the user requests a pause, it can only stop at the current process state and cannot be changed to a pure copy deliverable.
After batch generation has been confirmed, do not ask again whether the scope, title, body text, or script should be delivered first; directly generate the batch images according to the confirmed Brief, and when there is a failure, only report the specific failed pages and the downgrade/retry options.

Progressive loading

Only read the references needed for the current task:

Currently needed	Read
Determine triggers, inputs and outputs, and prohibitions	`references/skill-contract.md`
Automatically research reader needs, content opportunities, evidence, and competitive angles	`references/domain-research-brief.md`、`references/marketing-decision-model.md`
User decision order and the pause point at each step	`references/human-decision-first-workflow.md`
Vague intent, multiple decision popups, cover MVP, and parallel production after confirmation	`references/interactive-mvp-decision-system.md`
Remove AI flavor, title fulfillment, cover text quality	`references/writing-and-cover-quality.md`
Platform risk, commercial follow-up, boundaries of real materials	`references/platform-risk-boundaries.md`
Cover and page-by-page composition, text placement, visual tasks, page rhythm	`references/page-art-direction-system.md`
Visual route, realism, small-image readability	`references/visual-system.md`
Image2 image Brief, prompts, full image-text package	`references/image-generation-workflow.md`
Optional canvas/visual workbench slots, acceptance, annotation rework, and page-by-page asset list	`references/visual-workspace-contract.md`; after confirming that a specific provider is available, read the corresponding adaptation list
Optional visual workbench inquiry status, deduplication, startup, and status readback evidence	`contracts/visual-workspace-session.schema.json`、`evals/visual-workspace-offer.eval.json`
Failure modes, acceptance, and review	`references/failure-modes.md`、`references/evaluation-method.md`

When real details are insufficient, use worksheets/real-detail-intake.md. When a fillable workbench is needed, use assets/rednote-note-workbench.html。

Complete output structure

The workbench file for a complete note is output in a 15-section structure; for section names, order, and field definitions, see references/evaluation-method.md(authoritative source; this file does not repeat the list to avoid drift between two places). Some tasks output only the relevant sections, but it must be clearly stated which content is not within this scope. The final chat reply should not use these 15 section titles by default; the final chat uses the "Final Chat Delivery Structure" below.

Final Chat Delivery Structure

After the complete image-text is finished, deliver in the chat in the following order, without outputting the full internal workbench content:

完成摘要: whether native selection is used, whether the cover and inner pages are generated, and the validation results.
封面预览: display or list the newly generated covers for this round.
发布标题首选。
标题备选: at least 8.
正文草稿: copy-ready publishing version.
承接方式: Public interaction boundaries and low-risk responses when proactively asked.
标签: Mainly 3-6.
图片清单: Newly generated cover and 6-8 inner pages for this round; if blocked, write the blocking reason.
归档文件: Full 15-section workspace file path, as supplementary evidence.

Must-check items

The title and cover promise must be fulfilled by the body text and image pages.
A complete note must provide at least 8 title alternatives.
A complete note must provide at least 3 sets of cover text alternatives.
The body text must fulfill the marketing route selected upstream: for the trust/recommendation route, check real scenarios, process evidence, and restrained boundaries; for the search/save route, check methods, steps, common mistakes, and saveable conclusions; for the comparison/consultation route, check judgment criteria, suitable/unsuitable, and evidence; for the opinion/discussion route, check clear judgments, examples or counterexamples, and discussion boundaries. Do not write "scenario—turn—closure" as a fixed template for all body text.
Image prompts must distinguish between cover and inner pages, and include subject, composition, lens/perspective, lighting/material, and text treatment.
Cover visual verification must clearly state the route, evidence, readability in the mobile feed, and confirmation status.
Batch image production must state whether it is final, partially complete, or blocked, and write the reason and evidence.
The final chat delivery must not contain a full 15-section workspace stack; if the full structure is needed, write it to an archive file and give the path in chat.

Script

Run only when the input file exists:

node scripts/check-rednote-note.mjs <output.md>
node scripts/check-rednote-chat-output.mjs <final-chat-output.md>
node scripts/check-chat-output-fixtures.mjs
node scripts/check-page-direction-fixtures.mjs
node scripts/check-narrative-route-fixtures.mjs
node scripts/check-visual-workspace-contract.mjs
node scripts/build-rednote-visual-pack.mjs <output.md> <output-dir>
node scripts/build-rednote-realshot-handoff.mjs <output.md> <output-dir>
node scripts/build-rednote-final-replacement-pack.mjs <output.md> <output-dir>
# 把已生成封面/内页包成手机信息流测试页，用于缩略图可读性验收
node scripts/build-rednote-image2-feed-test.mjs <image-path> <caption> <output-dir>

If Image2 or another image generator has already produced the final image set, do not present SVG/static packages as parallel proof. Scripts and old packages passing only prove structure or historical results, not that images were generated in this round.

Local iteration boundaries

When this skill is not public or is still being polished privately, write failure retrospectives back into the skill itself, references, scripts, evals, or test-results; do not require "completing it within public evolution," and do not write internal test threads, private cases, or unpublished materials as release notes.

Final reply

The chat must display a copyable publishable version: selected title, title alternatives, body text, comment/private message boundaries, tags, image list or blocking reason, and missing real materials. File paths can only serve as supplementary evidence and cannot replace the body text results. The full 15-section workspace is placed in a file by default and is not expanded in full in the final chat.