Content Pipeline

Content Pipeline

AI-powered content pipeline: trend discovery -> deep research -> algorithm-optimized writing -> social publishing.

MCP Servers Required

Install via uvx (one-time, auto-cached):

uvx --from 'trend-pulse[mcp]' trend-pulse-server
uvx --from cf-browser-mcp cf-browser-mcp
uvx --from notebooklm-skill notebooklm-mcp

MCP Tools Reference

trend-pulse (12 tools)

Trend Data:

• get_trending(sources, geo, count): Query ALL 20 free sources. sources="" for all. geo: "TW"/"US"/"JP"/"". count: 20.
  • Sources: google_trends, hackernews, mastodon, bluesky, wikipedia, github, pypi, google_news, lobsters, devto, npm, reddit, coingecko, dockerhub, stackoverflow, producthunt, arxiv, lemmy, dcard, ptt
• search_trends(query, sources, geo): Cross-source keyword search.
• list_sources(): List all sources.
• take_snapshot(sources, geo, count): Save snapshot for velocity tracking.
• get_trend_history(keyword, days, source): Historical data with direction.

Content Guide:

• get_content_brief(topic): Writing brief with hook examples, patent strategies, CTA.
• get_scoring_guide(): 5-dimension patent scoring. Score >= 70 required.
• get_platform_specs(platform): Platform specs (char limits, algorithm signals, posting times).
• get_review_checklist(): Quality review checklist (7 checks).
• get_reel_guide(): Reels script guide (3 styles).

Search:

• search_threads_posts(query): Search Threads posts by heat score.

cf-browser (10 tools)

Headless Chrome via Cloudflare Browser Rendering. Use instead of WebFetch for JS-rendered pages.

• browser_markdown(url): Clean Markdown. Most used for deep research.
• browser_content(url): Full rendered HTML.
• browser_screenshot(url): Full page screenshot (PNG).
• browser_pdf(url): Generate PDF.
• browser_scrape(url, selector): CSS selector extraction.
• browser_json(url, schema): AI-driven structured data extraction.
• browser_links(url): Extract all hyperlinks.
• browser_a11y(url): Accessibility tree (low token cost).
• browser_crawl(url): Async multi-page crawl.
• browser_crawl_status(id): Check crawl progress.

notebooklm (13 tools)

Deep research + 9 downloadable artifact types (podcast, slides, report, quiz, flashcards, mindmap, datasheet, study_guide). ⚠️ infographic download is unreliable — use slides instead.

Management: create_notebook, list_notebooks, delete_notebook Research: add_source, ask, summarize, list_sources, research(mode="fast"|"thorough") Artifacts: generate_artifact(type, lang?), download_artifact(type, output_path) Pipelines: research_pipeline(sources, questions, output_format), trend_research(geo?, count?, platform?)

Video Synthesis (slides + podcast -> MP4):

1. Generate slides (PDF) + audio (podcast M4A) — sequentially
2. Download both
3. pdftoppm -png -r 300 slides.pdf slides_page
4. ffmpeg -framerate 1/$per_slide -pattern_type glob -i 'slides_page-*.png' -i audio.mp3 -c:v libx264 -pix_fmt yuv420p -c:a aac -shortest output.mp4

Visual Content Generation

Use NotebookLM to generate all visual content. It produces professional-quality slides and mind maps — far better than any alternative.

⛔ 嚴格禁止：不得使用 Pillow/PIL、ImageMagick、Python 程式碼或 HTML+Playwright 生成圖卡。 所有圖卡一律透過 NotebookLM nlm_generate(type="slides") 生成。

⚠️ Do NOT use infographic (download unreliable). Use slides for all visual content. ⚠️ NLM slides 多頁時 → 必須全部上傳 → 用 carousel_urls 發文（不要只發一張）。

Image Cards & Slides (圖卡 / 簡報)

Workflow: Create a NotebookLM notebook with the content as a text source, then generate the visual artifact.

1. create_notebook(title="Post Visual", text_sources=["<post content + key data points>"])
2. generate_artifact(name_or_id, "slides", lang="zh-TW")   → PDF slides (best for cards)
3. download_artifact(name_or_id, "slides", "downloads/card.pdf")

Choose artifact type by use case:

Carousel Posts (輪播貼文)

For multi-image carousel (Threads supports 2-20 images):

1. Create notebook with content organized as numbered sections (one per slide)
2. generate_artifact(name_or_id, "slides") → multi-page PDF
3. Download and split: pdftoppm -png -r 300 slides.pdf slide
4. Upload EVERY image — call upload_image for EACH slide PNG:

   upload_image(file_path="downloads/slide-1.png") → url1
   upload_image(file_path="downloads/slide-2.png") → url2
   upload_image(file_path="downloads/slide-3.png") → url3
   ...repeat for ALL slides
   

5. Publish as carousel — pass ALL URLs to publish_to_threads:

   publish_to_threads(text="...", carousel_urls=[url1, url2, url3, ...])
   

   Do NOT use image parameter — use carousel_urls when there are 2+ images.

⚠️ CRITICAL: When slides have multiple pages, you MUST use carousel_urls (not image). Single image parameter only posts the first image. carousel_urls posts ALL images as a swipeable carousel.

Video (影片)

Combine slides + podcast audio into MP4:

1. generate_artifact(name_or_id, "slides") → PDF
2. generate_artifact(name_or_id, "podcast") → M4A audio (run sequentially, NOT parallel)
3. Download both artifacts
4. pdftoppm -png -r 300 slides.pdf slides_page
5. ffmpeg -framerate 1/$per_slide -pattern_type glob -i 'slides_page-*.png' -i audio.mp3 -c:v libx264 -pix_fmt yuv420p -c:a aac -shortest output.mp4

Tips for Better Visuals

• Language: Always pass lang="zh-TW" for Traditional Chinese content
• Rich text sources: The more context you feed into the notebook, the better the visual output
• One topic per notebook: Don't mix unrelated topics — create separate notebooks
• Add URLs as sources: add_source(name_or_id, url="...") for reference material — NotebookLM will incorporate key data into visuals

Mandatory Rules

0. Workspace Containment

All files MUST be saved within the session workspace directory. Use relative paths like downloads/card.pdf. NEVER use ~/Downloads, ~/Desktop, or any absolute path outside the workspace.

1. Read Original Sources

NEVER write content based on titles/metadata alone.

• Single topic: read >= 1 primary source via browser_markdown(url)
• Controversial: read >= 2 sources (both sides)
• Data claims: find original data source

2. Timeline Verification

Every fact must have a verified timestamp. Discard anything > 48 hours old.

3. Use ALL Sources

get_trending with sources="" to query ALL 20 sources. Do NOT filter unless user explicitly asks.

Meta Patent-Based Scoring (5 Dimensions)

Quality Gates (ALL must pass before publishing):

• Overall Score >= 70
• Conversation Durability >= 55
• Hook: 10-45 chars with number or contrast
• CTA: clear question or poll
• Timeline: all time words verified
• Source: every claim traceable
• No AI filler words

Pipeline Workflow

1. Discover: get_trending(sources="", geo, count=20)
2. Read Source: browser_markdown(url) — MANDATORY
3. Verify Timeline: Check dates, discard stale
4. Research: browser_markdown + WebSearch, 2-3 sources min
5. Brief: get_content_brief(topic)
6. Create: Write -> patent check (5 dimensions)
7. Score: get_scoring_guide — self-score (>= 70)
8. Review: get_review_checklist — final check
9. NotebookLM (optional): create notebook -> generate artifacts
10. Publish: provide ready-to-post content or use publish tool