lov-article-creator

Genius Writer · Genius Writer

Turn scattered facts, research materials, or existing drafts into a deliverable WeChat Official Account article package, not just a piece of Markdown.

Each complete delivery simultaneously includes: a fact-constrained body, a topic-adapted editorial template, body illustrations when the topic requires them, a branded horizontal cover, an independent 4:3 horizontal body lead image, source and publishing metadata, and a reviewable acceptance report.

Triggers

Activate when

"Write a WeChat Official Account article in my writing style, and prepare the cover and body lead image."
"Turn these research materials into a uniformly branded WeChat Official Account long-form article."
“Create a branded WeChat article package from these notes.”
“Turn this draft into a publish-ready WeChat article with a wide cover and a 4:3 body hero.”
"Turn this existing Markdown into a complete branded WeChat Official Account version."
"Faithfully repost this partner article, keeping the original text and adding our opening and closing."

Do not activate when

Only asked to rewrite a passage or match a personal tone, without needing a WeChat Official Account template and visual package; use the writing style capability.
Only asked to read, modify, or verify an existing draft in the WeChat Official Account backend; use the lov-publish-wechat-article existing-draft pipeline.
Only asked to write an existing file into the draft box or formally publish it; use the WeChat Official Account publishing capability.
Only asked to generate a single isolated image, without needing article-level proposition, branding, and dual-ratio acceptance.

Product contract

One article, one package. The body, sources, cover, lead image, and acceptance record are located in the same article directory.
Facts before voice. First establish a fact ledger, then apply the writing style; do not use the first person to fabricate experiences, numbers, or changes in judgment.
Objective before outline. First clarify the creative conditions, motivation, communication goals, and desired_reader_change, then decide the structure. The proposition, opening, chapters, or draft given by the user are important constraints and evidence, and do not automatically equal an already sufficient and correct communication strategy.
One writing owner. New writing and rewriting must invoke lov-writing-style; it internally invokes the single anti-AI / authorship rule source lov-human-writing. This Skill only owns WeChat Official Account topic, structure, branding, and artifact rules, and does not copy general writing style rules.
Template is semantic. What is fixed is the information order, evidence thresholds, and brand components; it does not force all topics to use the same set of empty headings.
Heading roles stay separate. File names, platform titles, body H1, chapter H2/H3, and TOC labels are different artifacts. Chapter headings are responsible for dividing the reader's cognitive stages and gradually drawing them in, not for compressing paragraphs; the TOC only mirrors the final chapter structure and does not in reverse determine the body.
Brand comes from Profile. The publishing entity, Logo, official website, colors, and prohibited information are resolved from the shared Profile, and user private paths are not written into public articles or Skill source code.
Three image roles, not two. The complete pipeline must output 2.35:1 share cover and an independent 4:3 horizontal body lead image; the finished share cover must not be directly used as the body lead image. The third role is body illustrations: whether they are necessary is determined by topic, and the determination result must be explicitly recorded; it cannot default to zero just because no one asked.
Body figures are decided, never defaulted. Every complete pipeline must perform one body illustration determination and write it into the manifest; when determined as "needed", hand it to lov-illustrate for production, and when determined as "not needed", record the reason. The absence of a determination is itself a defect.
Evidence figures outrank decorative ones. The priority for body images is real product screenshots, traceable data charts, programmatically synthesized images, and only last AI concept images; for articles whose argument is carried by data, the body must not use AI concept images to impersonate evidence.
Benchmark claims require benchmark evidence. Once an article provides rankings, radar charts, or quantified pros and cons, it must disclose before the results the test subjects, inputs, complete Prompt, execution environment, evaluation metrics, scoring rules, at least one item-by-item scoring example, and reproduction method; scores must not appear before the method.
Publication identity wins. The WeChat Official Account publishing entity is separate from the parent brand; the cover uses only the publishing entity's official white horizontal lockup, and the square icon, parent brand Logo, and orange variants cannot substitute for it.
Generated is not accepted. The existence of a file does not equal completion; it must pass the size, citation, fact, brand, and mobile reading quality gates.
Publishing is separate. By default, stop at a locally publishable article package; writing to the draft box or formal publishing requires explicit authorization and is handed off to downstream publishing capabilities.
Platform title is not body content. Canonical Markdown retains a unique H1 as the single source of truth for the title; when writing to WeChat Official Account, the title goes into the platform title field, and the body hides this H1 by default to avoid a repeated title with the same name appearing before the first image.
References stay quiet. Source materials, supplementary links, and further reading interspersed in the body use low-contrast small italic source notes by default, and do not compete for visual hierarchy with section headings, body arguments, or primary action entry points.
Repost source is frozen. When reposting, the source body is a verbatim frozen zone; authorial optimization applies only to newly added blocks by the publisher, and copyrightMode is fixed as reprint。

User Profile

Each run first reads skill.yaml as declared by user-profile/v1 context, resolved in the order of the current request, project facts, this Skill records, shared preferences, brand and user Profile, and security defaults.

Templates, writing styles, brands, and cover ratios that the user directly declares and wants to use long-term are written via scripts/profile_store.py record --confirm into skills.lov-article-creator.records. Public brand facts use brand.*; inferred values, credentials, and internal materials must not be persisted.

Skill Kit Modules

Before full execution, read kit.yaml and the selected module:

$SKILL_DIR/skills/article-writing/SKILL.md — fact ledger, topic structure, and the body writing handoff for lov-writing-style .
$SKILL_DIR/skills/editorial-template/SKILL.md — fixed WeChat Official Account template, brand components, and article package metadata.
$SKILL_DIR/skills/body-figures/SKILL.md — body image decision gate, material priority, and the output handoff for lov-illustrate .
$SKILL_DIR/skills/cover-package/SKILL.md — art direction, horizontal cover, 4:3 body first image, and deterministic Logo composition.
$SKILL_DIR/skills/quality-gate/SKILL.md — acceptance of content, writing style, brand, images, and package structure.

Modules are hard dependencies of this Kit and are distributed with the source code. External adjacent Skills are handed off only through explicit files and are not hidden dependencies.

Pipelines

Pipeline	Applicable result	Module
`full`	Generate a complete article package from materials	writing → template → figures → cover → quality
`rewrite`	Rework an existing draft into a unified WeChat Official Account version	writing → template → figures → quality
`visual`	Body is finalized, only supplement the dual-ratio visual package	cover → quality
`figures`	Body is finalized, only supplement the body images	figures → quality
`brand`	Complete the structure, brand, cover description, and dual-ratio visuals for an existing local article	template → figures → cover → quality
`repost`	Freeze the source body, and only add the publisher's opening, source attribution, closing, and brand section	template → cover → repost audit → quality
`audit`	Only check the existing article package	quality

When the user says "write a WeChat Official Account article" and does not explicitly narrow the scope, use by default full。

Workflow (MANDATORY)

Step 0: Resolve the Kit and durable defaults

Locate the Skill root directory, and check kit.yaml, required modules, references, scripts, and assets.
Read the shared Profile and skills.lov-article-creator.records。
Separately parse the parent brand and the publishing entity of the official account, and read the publishing entity's official white horizontal logo, brand website, writing style, aspect ratio, and output directory; do not fall back to the parent brand, a square logo, or an orange logo.
When missing information would change the user-visible final product, ask only one focused question; use safe defaults for everything else and continue.

Step 1: Build the truth ledger

Fully read the input files, current project materials, images, and the user's explicit judgments, and internally distinguish:

Facts, versions, numbers, quotations, and links that can be written directly;
Real experiences that can use the first person;
External claims that require source support;
Internal context that only aids judgment and must not be made public;
Evidence gaps that would prevent a reliable finished draft.

Compress the old claim the article truly aims to correct, or the judgment it aims to advance, into a single proposition. When the input is insufficient to support a key conclusion, retain the boundary or request the minimum facts; do not fabricate a complete story.

Fully read Creative Strategy and Title System, and establish an internal strategy brief: currently available conditions, why the author wants to write at this moment, communication goals, the target reader's starting point, desired_reader_change, the desired reader action, existing evidence, and evidence gaps. Official account communication metrics can be opens, completion, likes, shares, comments, or experience, but conflicts, relationships, or results must not be fabricated for the sake of metrics.

Then determine which working mode this interaction belongs to:

Only a proposition or project materials: propose the proposition, reader change, and structure top-down;
An opening, chapters, or explicit writing approach has already been given: treat them as accepted constraints, fill the gaps, and do not silently rewrite;
Voice transcripts, thoughts, or material piles: extract facts, motivations, and structure bottom-up;
A complete draft already exists: iterate with minimal changes, first preserving the parts the user has proofread, then fixing the paragraphs that truly block the goal;
Mixed input: fix the levels the user has already decided, and use the corresponding method only for the undecided parts.

If a plan provided by the user conflicts with the real goal, neither blindly follow it nor arbitrarily override it. First point out the specific conflict, the missing evidence, and their impact on the communication goal, then give one executable alternative; only ask the user to decide if this choice would substantially change the final product.

At the same time, establish reader contract: the publishing channel, target readers, what readers already know when they open the article, and the boundaries of the final product. The final draft for an official account by default faces public readers who did not participate in the conversation between the author and the Agent. Previous draft rounds, user criticism, revision notes, and task process belong to internal context; unless they themselves are the subject of study and are fully established within the main text, they must not enter the article's proposition or introduction.

Step 2: Write for meaning

Read the article writing module, and hand the truth ledger, reader contract, subject matter, and publishing channel to lov-writing-style。lov-writing-style for personal writing style, and internally invoke lov-human-writing to perform an authorial audit; do not bypass this chain and write another set of 'de-AI-ify' rules.

Continue executing the official account-specific structure:

A real conflict, scene, result, or core judgment appears within the first 300 characters.
Give the title and the first 300 characters separately to zero-session-context as a cold reader; all versions, characters, events, and references must have antecedents in the visible final product. If dangling phrases such as 'the previous version / the last draft / this rewrite / according to your request' appear, rewrite directly; they must not enter the next stage.
The main text follows desired_reader_change Choose the cognitive path; definition, distinction, evidence, cost, and action are merely available modules, not a fixed order.
Each major section has at least one fact, case, number, personal experience, or source.
Long sentences carry explanation, short sentences deliver the hammer blow; keep alternating between one-sentence paragraphs and long paragraphs on mobile.
Complete the facts and argument first, then adapt to personal style; do not copy catchphrases, do not manufacture false emotion.
First determine the subject matter. Research, investigation, comparative testing, benchmark, or articles with rankings/radar charts use the paper-style methods structure:调研对象、测试方法、Prompt、评价指标、评分方法、评分示例、复现方法、测试结果、局限性、结论. These titles prioritize retrieval and reproducibility, stay plain, and are not forcibly changed into suspense sentences or judgment sentences.
Research and evaluation articles must, before 测试结果 , let readers see: the version or hash under test, unified input, the actual execution Prompt, model and key parameters, isolation conditions, sample count and repetition count, blind evaluation or randomization method, metric definitions and weights, item-by-item scoring examples, runnable commands, and the raw artifact index. If any one is missing, you can only write "exploratory observation" and cannot give a definitive ranking.
First determine the cognitive stage of the section, then name the H2/H3. Use orthogonal concepts for parallel issues; do not treat character gender, "Case One / Case Two," or accidental sequence as essential; when there is no real order, do not use "first... then..." to manufacture a process. When the user has explicitly given section names, keep them by default unless they clearly conflict with the body text or communication goal.

repost The pipeline does not rewrite the source body text. Read in full source fidelity、repost increment and authorship boundary, place the source into the sole data-repost-source="true" frozen zone; only hand the newly added opening, source attribution, closing, and brand microcopy to the writing chain.

Step 3: Apply the editorial template

Read $SKILL_DIR/references/article-template.md and the editing template module, and fit the body text into a fixed semantic structure:

Title and publishing metadata;
4:3 horizontal body lead image;
lead;
body text and a stable H2/H3 hierarchy; Lovpen WeChat HTML can generate a TOC from this, while canonical Markdown does not hard-code a presentational table of contents by default;
conclusion or "written at the end";
when there is a real destination, add code, original text, blog, or further reading;
stable brand endnote;
cover note after the conclusion and before the brand endnote: for real artworks write "this issue's cover"; for generative covers write a copyable "cover Prompt";
read_original_url, source, and publishing fields.

Use scripts/build_article_package.py to build the table of contents and manifest. Do not create nonexistent links, products, or author experiences just to fill the template.

When an existing article uses the brand pipeline, additionally read branded version。publication、brand and products must be kept separate; the brand endnote does not enumerate products by default, and only includes them when directly related to the body text or explicitly requested by the user.

Step 4: Decide and produce body figures

Read body illustrations, perform one explicit determination on the assembled article.md , and write the result into the manifest's body_figures：

Determine whether this article needs body illustrations. Data research, evaluation, comparison, tutorial, and process articles default toneeded: the argument of such articles is carried by curves, distributions, comparisons, or steps, and readers cannot verify the conclusion without seeing the charts. Opinions, essays, short reviews, and announcements default tonot needed。
When judged as "not needed", write the reason into the manifest and skip to the next step; do not skip silently, and do not force-add images just to make up the count.

When judged as "needed", hand over the in-package article.md to lov-illustrate for planning and production:

# 在文章包目录内执行，让图片落在 figures/ 并按相对路径回写正文
lov-illustrate article.md --style <brand-style> --max <n>

The material priority is fixed as: real product screenshots, traceable data charts, programmatically synthesized images, and only last AI concept images.The core numbers of the article must be carried by data charts; AI concept images cannot be used to replace evidence.
Images land in the in-package figures/, and the body references them by relative path; then rerun scripts/build_article_package.py to make body_image_paths consistent with the actual files.
repost The pipeline does not apply to this step: the source body is a verbatim frozen zone, and publisher images must not be inserted.

The share cover and the 4:3 body lead image are handled separately by Step 5, do not belong to this step, and cannot be used as body illustrations.

Step 5: Produce the cover package

Read $SKILL_DIR/references/cover-system.md、$SKILL_DIR/prompts/cover-art.md and the cover module:

Extract one main visual metaphor from the article's proposition; do not use robots, code UI, neon, tech particles, or pseudo-text to fake an "AI feel".
Generate an art base image containing no Logo, title, letters, numbers, or watermark, reserving safe zones for the share cover and the body lead image respectively.
Use the same visual language to compose the share cover and the 4:3 body lead image separately; do not reuse the finished share cover with the Logo already overlaid, and do not mechanically stretch.
Run scripts/compose_covers.py, deterministically composite using the publishing entity's official white horizontal raster Logo, and output PNG, JPG, and manifest; the script must reject square and non-white Logos.
The share cover is used for the WeChat official account message list;4:3 the body lead image is located at the first block of the body, before the lead paragraph.

Step 6: Run the quality gate

Read $SKILL_DIR/references/quality-gate.md, and run:

python3 "$SKILL_DIR/scripts/validate_article_package.py" \
  --package ARTICLE_PACKAGE \
  --json

At the same time, perform a manual semantic check: facts and quotations, author position, writing style and rhythm, title and summary, further reading, brand entity, cover focus, Logo integrity, and the mobile reading effect of the body lead image. Research and evaluation articles must also check item by item the experimental input, Prompt, environment, scoring rules, raw results, and reproduction entry point.audit only reports; after other pipelines find problems, fix them directly and rerun.

The repost package must also run:

python3 "$SKILL_DIR/scripts/audit_repost.py" \
  --source-text SOURCE_TEXT \
  --edition-html EDITION_HTML \
  --source-account SOURCE_ACCOUNT \
  --source-url SOURCE_URL \
  --copyright-mode reprint

Step 7: Deliver the package

By default, output to the current workspace output/articles/<slug>/：

article.md
article-manifest.json
sources.md
cover/
  art-master.png
  wechat-cover-wide.png
  wechat-cover-wide.jpg
  article-opening-4x3.png
  article-opening-4x3.jpg
  cover-manifest.json
quality-report.json

repost The pipeline additionally outputs the source snapshot,source-assets/、edition-manifest.json and the fidelity audit receipt; the remote Lovpen HTML and publishing receipt are generated by the Publisher.

Report the fact support scope, actual files, Profile sources used, and verification results. Without authorization, do not create WeChat official account drafts, do not formally publish, and do not send external messages.

Dependencies

Python 3.9+、PyYAML、Pillow。
A complete cover requires the host to provide image generation or legal material input; deterministic compositing itself does not require the network.
lov-writing-style is the sole external owner of the body writing style; its authorship audit is completed built-in by lov-human-writing .
lov-illustrate is the sole external owner of body illustrations; this Skill is only responsible for judging whether illustrations are needed, handing over article.md, and accepting the written-back results.
Publishing and WeChat official account backend editing belong to the downstream lov-publish-wechat-article, and are not part of this Skill's default delivery.

References

Article style acceptance must retain writing-review.json, and verify the final text hash and paragraph-by-paragraph semantic review record according to references/quality-gate.md . Statistical pass does not mean style pass; packages with missing records must not be marked as prepared.