Mhhya youtube ai to wechat

YouTube AI To WeChat

Turn a YouTube AI topic or video into a publishable WeChat draft, with公众号 readability as the default priority and source fidelity used as a factual constraint rather than the main writing goal.

Workflow

0. Verify dependencies first

Before transcript extraction or draft push, run:

bash scripts/ensure_dependencies.sh

This script checks that these sibling skills exist under the same skills/ root:

• baoyu-youtube-transcript
• baoyu-post-to-wechat

If one is missing, the script should try to install it with clawhub install <slug> --dir <skills-root>. If auto-install fails, stop early and report exactly which dependency is missing instead of continuing with a half-broken pipeline.

1. Lock the job shape

Classify the request into one of three entry points:

1. Discovery only: user wants recent/high-heat AI videos or topics from YouTube
2. Video-to-article: user has already chosen a specific video
3. Full pipeline: user wants discovery, drafting, images, and push to WeChat drafts

If the user already chose the video, skip discovery.

Default stance:

• use the selected video as the factual base
• write the finished article as a standalone公众号 piece
• write like a person noticing a concrete work change, not like a neutral industry-summary machine
• do not over-prioritize source-faithful recap style unless the user explicitly asks for it
• avoid repeatedly writing phrases like “视频中提到 / 视频里说 / 这条视频讲了什么” in the finished article
• do not append a “来源于某视频” footer unless the user explicitly wants attribution shown in the article body

2. Scan YouTube heat

When discovery is needed:

• search recent YouTube AI content with recency and heat signals
• use a tiered time window unless the user says otherwise:
  • 热点类 / 强时效话题: prefer the last 7 days
  • 深度类 / 解释型 / 工作流型话题: prefer the last 30 days
• extract practical signals: title, channel, rough view count, recency, why it is hot, and which angle fits a公众号 article
• group findings into 3–5 clusters instead of dumping a long raw list

Good cluster examples:

• model/company drama
• coding agents / Claude Code / AI workflow
• AI jobs / social impact
• open source tools / agents
• Gemini / OpenAI / Anthropic product shifts

3. Confirm the writing mode

For the chosen video, decide the writing mode before drafting:

• 公众号版: default; keep the core facts but adapt rhythm, hook, and readability
• Hybrid: preserve the main structure while adding a small amount of official supporting context
• Source-faithful: only when the user explicitly asks for tighter alignment with the video

Default to 公众号版.

Before drafting or revising when voice/tone or SEO matters, read:

• references/style-profile.md
• references/edit-playbook.md
• references/seo-profile.md

4. Extract transcript

Before this step, make sure bash scripts/ensure_dependencies.sh has passed.

Read and follow the sibling skill:

• ../baoyu-youtube-transcript/SKILL.md

Use it to fetch transcript, cover, and metadata for the selected video. Reuse its generated files instead of re-transcribing manually.

5. Draft the article

Write the article into:

• post-to-wechat/YYYY-MM-DD/<slug>/article.md
• post-to-wechat/YYYY-MM-DD/<slug>/image-plan.md

image-plan.md is now a default companion file for article drafting unless the user explicitly says they only want text and do not need body-image planning.

Minimum frontmatter:

---
title: 中文标题
description: 一句话摘要
coverImage: imgs/cover.jpg
---

Drafting rules:

• before drafting, lock 3 anchors for yourself: one interrupted work scene, one slightly sharper human judgement, and one concrete question the正文 will answer
• before drafting正文, also lock one single-sentence main judgement that the whole article will serve; if you cannot say the whole piece in one sharp sentence, you are not ready to draft
• before drafting正文, lock a judgement-serving section spine with only 3–4 sections; for each section, answer two things before you write it: (a) what exact job this section does for the main judgement — choose one from scene / action chain / friction or contradiction / consequence or watchpoint; (b) what exact proof asset would sit after it. If a section can only be described as “继续解释主判断”, or two adjacent sections are doing the same job, the spine is not ready yet
• before drafting正文, also lock a proof-asset spine in image-plan.md: 1 cover source + 2–4 body-image slots, each bound to a specific section/paragraph beat with at least these fields: slot / section anchor / proof point / why this image belongs here / screenshot target / source route / fallback
• before you call article.md finished, convert every kept body-image slot into a real inline Markdown image node inside the正文 at its bound paragraph beat. Default shape: paragraph → ![<caption cue>](imgs/body-1.jpg) → next paragraph. Use the same slot numbering / filenames as image-plan.md so the publish chain can read and upload them.
• when image planning matters, read references/image-planning.md before you fill image-plan.md
• write each body-image target as a screenshot sentence: say which page, which block, which UI region, which doc position, which changelog section, which repo area, or which before/after comparison the editor should go capture. 官方 docs / 官方产品页 / GitHub / 新闻页 by itself is not a finished target.
• also lock a light SEO pack from references/seo-profile.md: 1 core search phrase, 2–3 support phrases, and 3–5 tags
• lock one literal primary search object before drafting; unless the user explicitly wants a different title strategy, the main title or description should carry that phrase and the first screen should echo it naturally once; do not let the object drift across neighboring labels like CRM / Agentforce / 销售平台 or 平台层 / 上下文管理 / 工具路由
• follow references/style-profile.md for the default public title/opening/body/persona defaults in this repo
• if the main title feels merely correct, too abstract, or too much like an update summary, read references/title-patterns.md before you lock it
• keep the Chinese title aligned with the source video when the user asks for consistency
• preserve the original video’s key claims, examples, and distinctions
• optimize for公众号 readability: short paragraphs, strong sectioning, direct wording
• keep SEO human-first: make the title/description/body easier to search without turning them into keyword stuffing
• for company / product / open-source shift topics, make the main title carry both: (a) one concrete change object from the source, and (b) one plain-language reader consequence or workplace effect in the same line; when possible, lead with the consequence/shift first and avoid front-loading dense term lists
• for abstract platform / infrastructure topics, do not let a raw phrase like AI 平台层 / 平台能力 be the main title object by itself; first translate it into one workplace-layer object the reader can picture and search, then reuse that same object in the title or description and the first screen before you bring in technical nouns like identity / context / tool routing
• keep the opening closer to 140–170 Chinese characters when possible; for company / product / open-source shift topics, use a reader-consequence lead: start from what changes for the reader, team, or workflow first, then name the source move, and end the first screen with one explicit question/promise instead of a generic trend summary; for open-source / toolchain topics, make that promise concrete by naming the exact work or adoption question the rest of the piece will answer
• when the source supports it, open from one specific changed work moment, role, or action chain; do not default to generic openings like “如果你还以为… / 表面上看… / 这次更新最值得看的是…”
• keep at least 2 concrete work scenes or action chains in正文 when the source supports it; name role, tool, action, and what step gets saved or taken over, instead of only talking about “效率 / 组织 / 工作流 / 趋势”
• make at least one middle section enter from friction / contradiction / exception / failed handoff / boundary, not from another smooth restatement of the thesis
• if a sentence could fit another company or video by swapping nouns, rewrite it more specifically
• do not lean on repeated stock pivots such as “不是 X，而是 Y / 这意味着什么 / 换句话说 / 这并不意味着 / 表面上看”; use at most one if it is genuinely doing work
• vary pacing: not every paragraph needs to be a fully wrapped mini-summary; mix a longer scene paragraph, a shorter punch line, and 1–2 controlled human-sounding judgements
• do not spend multiple sections restating the same thesis with near-synonyms; each section should add a new layer such as scene, friction, who gets squeezed first, or what the reader should watch next
• if the last section only sounds like a calmer re-summary of the same judgement, cut it or rewrite it into one sharper watchpoint / consequence / open tension
• avoid writing the article as a running transcript recap unless explicitly requested
• default to not exposing transcript/conversion traces in正文
• avoid repeated self-referential source phrases in正文
• omit “视频来源/内容来源于某视频” at the end unless the user explicitly wants it
• use smoother Chinese phrasing instead of literal subtitle translation
• when in tighter source-faithful mode, preserve claims more strictly but still keep the article readable
• when useful, generate a small seo-notes.md beside the article with the main title, alternate titles, description, core phrase, support phrases, and tags
• if a draft feels generic, too smooth, or “AI-written,” run the revision moves in references/edit-playbook.md before delivery

6. Add images

Prefer source-aligned visuals in this order:

1. video cover
2. official source pages or official docs screenshots
3. official product/news/images tied to the same topic
4. video-frame screenshots only if the user asks for them or no better official image exists

Image rules:

• keep each image tied to the section it supports
• prefer 2–5 body images, not a gallery dump
• store images under post-to-wechat/YYYY-MM-DD/<slug>/imgs/
• image-plan.md should default to at least: one cover + 2 body images; body images should support different section beats, not pile onto the same paragraph unless the source really demands it
• for every body image, state: which section it follows, what judgement it proves, what exact screenshot object to capture, and where the editor should go find it first
• every kept body-image slot must also appear inside article.md as a matching inline Markdown image reference such as ![图注](imgs/body-1.jpg) placed right after the paragraph or section beat it proves; planning it but not inserting it into the article does not count as done
• if the plan only says 官方 docs 页面 / 官方产品页 / 官方新闻页 / GitHub / changelog without naming the exact target block, the image plan is still unfinished
• for abstract platform topics, do not escape back to “无图可配”; translate the claim into a screenshotable carrier such as a permissions panel, connector list, workflow graph, admin setting, changelog block, benchmark table, repo tree, or before/after comparison
• do not treat images as decoration; every planned image should help the reader see one concrete change, handoff, boundary, or evidence point named in正文
• drafting is not complete until image-plan.md contains screenshot-level targets that another editor could execute without re-guessing your intent
• drafting is also not complete until article.md itself contains at least 1–2 inline body-image Markdown references beyond the cover frontmatter, and those references can be matched back to the kept slots in image-plan.md

7. Push to WeChat drafts

Before this step, make sure bash scripts/ensure_dependencies.sh has passed.

Read and follow the sibling skill:

• ../baoyu-post-to-wechat/SKILL.md

Publishing rules:

• prefer API mode when credentials are configured
• fall back to browser mode when API credentials are missing or the user prefers browser
• if API fails with IP whitelist errors, tell the user the exact outbound IP and ask them to whitelist it
• publish the local markdown file directly; let the posting skill handle conversion

8. Report back clearly

Always tell the user:

• chosen angle or video
• article file path
• image-plan path
• writing mode used
• what images were used or planned for which sections
• whether draft push succeeded
• WeChat draft media_id if API publish succeeded

Read on demand

Also use:

• scripts/ensure_dependencies.sh to verify/install sibling dependencies before transcript extraction or publish

Read these references on demand:

• references/workflow.md for the full SOP, anti-AI voice guardrails, revision checklist, file conventions, and handoff format
• references/style-profile.md for the account’s stable writing profile
• references/edit-playbook.md for recurring user-feedback fixes and revision moves
• references/seo-profile.md for light, human-first公众号 SEO defaults
• references/image-planning.md for screenshot-level body-image planning, topic-specific asset presets, and executable handoff rules