Publish Md To Wechat

publish-md-to-wechat

Publishes Markdown articles to WeChat Official Account drafts with professional visual styling, automatic image handling, and cover generation.

Operating Modes (strict boundary)

Mode A — Article Publish (公众号图文)

Use when user wants Markdown -> WeChat draft publishing.

• Inputs: --md, style, optional thumb/title
• Output: WeChat draft media id
• Command family: scripts/wechat_publisher.py

Mode B — Video Render (视频号)

Use when user wants vertical MP4 generation.

• Inputs: source article plus duration, tone, audience, visual style, and optional Edge TTS voice
• Output: local .mp4
• Command family: scripts/video_publisher.py
• Important: The agent plans the video and creates the project files; the script renders, narrates, subtitles, composes, and captures validation frames.

Mode C — Translation & Repost (全文翻译/转载)

Use when user wants to translate a foreign-language article and publish it to WeChat.

• Trigger: "全文翻译"、"转载"、"translate"、"repost"、or source article is in English/other language.
• Inputs: source Markdown file (English or other language)
• Output: WeChat draft media id (same as Mode A)
• Workflow: Agent translates natively (no Python script) → saves to /tmp/translated_article.md → proceeds as Mode A.
• Important: Translation is done by the agent itself, not by any script. Do not invoke an LLM via Python for translation.

Do not mix Mode A and Mode B steps in one flow unless user explicitly asks for both deliverables. Mode C always ends with Mode A publishing — it is a pre-processing step, not a separate pipeline.

Phase 1: Environment & Credentials

Check dependencies:

ls .venv/bin/python 2>/dev/null || echo "MISSING"

If .venv is missing, run ./install.sh.

Credentials: Loaded automatically from .env in the current directory, or ~/.config/publish-md-to-wechat/.env as global fallback. Only ask the user for credentials if the script explicitly fails with an auth error.

Phase 1B: Mode C — Translation & Repost

Do not summarize or adapt — translate the full article faithfully.

1. Read the source Markdown file completely.
2. Translate the entire content to Chinese, preserving:
   • All Markdown structure (headings, lists, code blocks, blockquotes, images)
   • Original section order and hierarchy
   • Technical terms in English where standard (e.g. agent, token, context window, bash)
   • Code blocks verbatim — do not translate code
   • Author attribution in frontmatter: > 本文译自：[原文标题](url)，作者：[Author]（[Publication]）
3. Write YAML frontmatter with title, author, description in Chinese.
4. Save to /tmp/translated_article.md.
5. Proceed to Phase 2 with the translated file as --md.

Do not ask for confirmation before translating — if the user said "全文翻译/转载", just do it.

Phase 1C: Frontmatter Adaptation (Vault Articles)

Before publishing any Vault article, verify the frontmatter contains the fields WeChat publisher needs to render the article header correctly.

The publisher renders a metadata block at the top of every article using these specific keys:

Script fallbacks (built-in): The publisher script automatically falls back:

• published → date if published is absent
• description → excerpt if description is absent

So these alias pairs do not need to be manually patched before publishing. The remaining three (author, tags, source) have no automatic alias — add them to frontmatter if missing.

Minimal complete frontmatter for a Vault → WeChat article:

---
title: 文章标题
slug: type-YYYY-MM-DD-topic          # ai-dive API
date: YYYY-MM-DD                     # ai-dive API
published: YYYY-MM-DD                # WeChat header display
type: analysis                       # ai-dive API
author: Rafa
tags: [标签1,标签2, 标签3]           # WeChat header display
excerpt: 摘要文字                     # ai-dive API
description: 摘要文字                 # WeChat header display (same content as excerpt)
source: AI-DIVE · ai.air7.fun       # WeChat header display
og_image: https://...               # ai-dive cover (fill after upload)
cover: assets/topic-cover.png       # local preview
---

Note: date and published carry the same value; excerpt and description carry the same content — the two systems use different keys for the same concept.

Phase 2: Read the Article & Select a Style

Read the Markdown file to understand the content's tone. Then pick a style.

List all styles:

.venv/bin/python3 scripts/styles.py --list

Style selection guide:

Core trio quick guide:

• swiss — Technical articles, tutorials, reports, official announcements
• editorial — Opinion pieces, analysis, blog posts, personal brands
• ink — Humanities, culture, deep dives, literary content

Propose the style with a one-line reason. Skip if the user already specified.

Phase 3: Validate First (Recommended for complex articles)

.venv/bin/python3 scripts/wechat_publisher.py \
  --md [PATH] --style [STYLE] --dry-run --out-html /tmp/preview.html
open /tmp/preview.html

Phase 4: Publish

.venv/bin/python3 scripts/wechat_publisher.py \
  --md [PATH] --style [STYLE]

Key flags:

• --thumb [PATH] — Custom cover (auto-generates if omitted)
• --title "[TITLE]" — Override title (auto-detects from frontmatter/H1)
• --no-verify-ssl — For SSL cert issues
• -v — Verbose logging

YAML Frontmatter: Title, author, description auto-extracted:

---
title: Article Title
author: Author Name
description: Short summary for WeChat article list.
---

Phase 4B: Generate WeChat Video (视频号)

By default, ALWAYS generate video WITH TTS voice unless explicitly told not to. Important: Before generating the video, you MUST ask the user for the following preferences if they haven't provided them explicitly in their request:

1. Target Duration (e.g., 30s, 60s, 90s)
2. Narration Tone (e.g., 专业克制, 轻松幽默, 热情洋溢)
3. Target Audience (e.g., AI开发者, 产品经理, 大众)
4. Visual Style (e.g., swiss, ink, minimal)
5. TTS Voice (optional; default zh-CN-YunxiNeural)

Do not guess or assume these 5 parameters. If the user simply says "convert to video", you MUST stop and ask them to specify duration, tone, audience, and style first.

Step 1: Agent Plans Scenes (DO NOT use Python scripts for LLM calls) Read the Markdown file. Act as an expert presentation editor. Natively create a video project directory containing scenes.txt and gen_scenes.py, based on the requested duration, tone, audience, and visual style.

Create the project first; do not start from a blank Python file:

.venv/bin/python3 scripts/video_publisher.py --init-project /tmp/video-project

Then edit /tmp/video-project/scenes.txt and /tmp/video-project/gen_scenes.py. The bundled generator is the visual baseline: retain its 1080×1920 canvas, brand line, page number, bottom bar, contrast, and common CSS; adapt the copy and layouts for the story.

CRITICAL: Narration Quality Rules (avoid filler/repetition)

• Audience/tone are global constraints, not per-scene slogans. Do not repeatedly say phrases like 面向AI开发者 / 关键在于 in every scene.
• Each scene narration must add new information (fact, example, contrast, action step), not template padding.
• Ban empty framing lines: 我们来看看, 关键在于, 总的来说 unless they introduce concrete content right after.
• Keep narration concise and spoken: typically 1-3 short sentences per scene; avoid paragraph-long monologues.
• Avoid repeating the same sentence pattern across scenes.
• Prefer concrete wording over abstract buzzwords.

Narration self-check before writing files:

1. Read all scene narrations end-to-end.
2. If any phrase stem repeats in 3+ scenes (e.g., 面向…, 关键在…), rewrite.
3. Ensure each scene has a unique takeaway in one line.
4. Remove all non-informational filler.

CRITICAL: HTML Scene Formatting for Vertical Video (9:16)

• scenes.txt contains one ordered scene per line: S1|spoken narration|optional written subtitle.
• gen_scenes.py must generate scenes/S1.html, scenes/S2.html, and so on at 1080×1920.
• Use a shared CSS system in the generator: stable brand line, page number, bottom information bar, high-contrast text, and a restrained accent color.
• The spoken and subtitle text can differ for URLs and English terms. For example, speak air7 点 fun but subtitle air7.fun.
• Keep each scene to one idea. The script derives timing from the generated audio, adds a short breathing interval, and burns subtitles into the video.

Step 2: Render and Compose (execution-only) Use the script to generate HTML, capture PNG scenes, synthesize narration, create subtitles, compose the MP4, and capture start/middle/end QA frames.

# 导出 MP4
.venv/bin/python3 scripts/video_publisher.py \
  --project-dir /tmp/video-project \
  --duration [SECONDS] \
  --voice zh-CN-YunxiNeural \
  --rate +6% \
  --out [OUTPUT.mp4]

依赖：ffmpeg、Python Playwright Chromium 和 edge-tts。

Step 3: Mandatory visual QA Open build/check_start.jpg, build/check_middle.jpg, and build/check_end.jpg. Check subtitle line breaks, readable contrast, overflowing copy, spelling, and narration-to-scene alignment. If any frame fails, modify only the affected scene in scenes.txt / gen_scenes.py, then rerun the renderer. Do not report a video as complete without passing this check.

Phase 5: Success Output

## 🚀 发布成功！

| 字段 | 值 |
|------|-----|
| **标题** | [Title] |
| **风格** | `[style]` — [one-line description] |
| **Media ID** | `[MEDIA_ID]` |

**下一步：** mp.weixin.qq.com → 草稿箱 → 预览 → 发布

Custom Styles

Style replication uses Playwright getComputedStyle() (not inline CSS parsing) for accurate colors, and automatically detects structural patterns:

• Heading style: bg-block (colored section bg) / left-border / underline / plain
• Blockquote style: left-border / full-box / plain

The renderer applies these structural hints when publishing with a custom style.

# Create from WeChat article (give it a meaningful name)
.venv/bin/python3 scripts/styles.py --url https://mp.weixin.qq.com/s/xxx --name custom-myname --no-verify-ssl

# List all styles (core / extend / custom)
.venv/bin/python3 scripts/styles.py --list

# Rename
.venv/bin/python3 scripts/styles.py --rename custom-old custom-new

# Use
.venv/bin/python3 scripts/wechat_publisher.py --md article.md --style custom-xxx

Troubleshooting

Known WeChat Rendering Constraints

These are platform-level limitations in WeChat's HTML renderer, handled automatically by the post-processor since v0.7.0: