公众号运营专家

先厘清要解决的问题

公众号运营的真实负担不在“不会写”，而在环节分散：选题靠拍脑袋，正文改到第五版还在调语气；排版套来套去容易破坏微信原生结构，预览又总在手机端翻车；发布前对医疗、金融、绝对化用语没底，标题 A/B 基本靠感觉，事后复盘也常常只看阅读量一个数字。

这套技能把操盘手式的工作方式封装成连续流程，目标是让你从没灵感一路走到可发布，并保留可追溯的产出。

它如何工作：能力与关键步骤

• 内容生成：爆款选题雷达根据行业关键词输出 10 个带情绪钩子的选题，并用热度标签标注预估空间；可接入微信搜一搜 / 新榜当日热点做简化 RSS 补充。正文阶段支持 /续写、/缩写、/改口语化、/加小标题 等行内指令，同时自动产出摘要、首图文案与文末引导语。
• 仿写模式：上传 3 篇历史 10w+ 文章后，提取口语化程度、金句密度、小标题风格与转折词偏好，先给文风特征表确认，再按指定语气仿写，减少风格漂移。
• 智能排版：首次采集主色、辅色、正文字号、行间距、头图模板与尾图配置，之后 AI 排版 直接套用 CSS，只叠加样式类，不改写微信编辑器 DOM。结构识别会把“3 个要点”转成卡片列表、“注意/警告”加底色框，并插入少量高频 SVG 动效；配图走国内合规生图 API，自动压缩并适配手机阅读。
• 发布前安全屋：敏感词实时标红，标题实验室一次给出 6 个标题并按情绪值、好奇心缺口、SEO 关键词密度打分，封面图按 2.35:1 主图与 1:1 次图自动裁剪。
• 复盘回流：发文后 24 小时做数据异常提醒，高打开率文章的关键词沉淀回选题库。

标准 SOP：定选题 → 写正文 → 智能排版 → 安全屋 → 发布/复盘。选题输出用表格（序号、选题、钩子类型、热度），正文默认 Markdown，排版产出保留原生结构的 HTML 片段，扫描结果标注位置与替换建议。

适用边界与注意点

• 外部热度、词库、数据必须标注来源与时效，不编造数值；分享率异常只是提示，不等于归因。
• 医疗、金融、绝对化用语只标红提示，不会替你放行违规表述；标题同样需要人工确认。
• 仿写只处理你有权使用的自有历史文章，不套用未授权第三方风格。
• 排版样式不能绕过平台限制，封面比例需严格遵守。

What it actually fixes

Running an Official Account is rarely blocked by writing alone. Ideation is guesswork, tone drifts across draft five, layout edits break WeChat’s native structure, compliance is a guess, titles go live without A/B, and post-publish reviews reduce everything to one vanity metric. This skill packages an operator-style workflow so you can move from no idea to publish-ready, with artifacts you can audit.

How it works: capabilities and key steps

• Content engine: topic radar takes industry keywords and returns 10 ideas with emotional hooks plus a heat label; a simplified RSS layer can pull WeChat SouYiSou / Newrank items. Drafting supports inline /continue, /shorten, /make casual, /add subheads, and auto-generates the summary, cover caption, and end CTA.
• Style cloning: after three of your historical 10w+ posts, it extracts colloquial level, density of punchy lines, subhead style, and transition-word preferences, shows the style sheet for confirmation, then writes in that voice without drifting.
• Layout engine: the first run collects primary/secondary colors, body size, line height, header template, and footer config. Later AI typeset applies CSS as style classes only, never rewriting the WeChat editor DOM. Structural recognition turns “3 takeaways” into a card list and flags note/warning blocks with backgrounds; only high-frequency SVG interactions are inserted. Illustrations come from compliant domestic image APIs, compressed and optimized for phones.
• Pre-publish safety room: sensitive words are flagged in place, the title lab returns six variants scored on emotional charge, curiosity gap, and SEO density, and cover images are auto-cropped to 2.35:1 hero and 1:1 secondary ratios.
• Review loop: a 24-hour post-publish check surfaces anomalies, and high-open-rate keywords flow back into the topic library.

Standard SOP: pick topic → draft body → typeset → safety room → publish/review. Topics are returned as a table (number, idea, hook type, heat), the draft stays Markdown, layout output is HTML fragments that preserve native structure, and scan results cite exact positions with replacement suggestions.

Fit and cautions

• External heat, word lists, and metrics must cite source and freshness; no fabricated numbers. An abnormal share rate is a prompt, not a diagnosis.
• Medical, financial, and absolute claims are flagged, not cleared. Titles still need a human decision.
• Style cloning uses only your own historical posts, not unauthorized third-party voices.
• Layout styling cannot bypass platform limits, and cover ratios are mandatory.