The Content Repurposing Pipeline: One Idea, Every Platform, One Agent
Every creator's advice is "repurpose everything." Nobody's advice includes actually doing it, because rewriting one idea five times is miserable. This pipeline makes the agent do the rewriting — five skills, each the best scorer in its slot.
Repurposing fails as manual work because each platform wants a different shape: articles want structure, decks want one idea per slide, LinkedIn wants a hook in the first line, WeChat wants a completely different register. The baoyu skill family covers each shape natively. Here's the pipeline in dependency order.
Stage 0 — capture: baoyu-youtube-transcript
Writing, 9.3/10. Clean structured transcripts out of any YouTube video — speaker turns, timestamps, paragraphs.
Why it's first: Everything downstream consumes this. Its output formatting is what makes the later stages good — the transcript arrives as readable sections, not a wall of captions.
Stage 0b — research: baoyu-url-to-markdown
9.2/10. Pulls any web page into clean Markdown.
Why: Repurposing isn't just format-shifting — it's enriching. One video + three reference articles becomes an article that's better than its source.
Stage 1 — the article: prompt-master
9.3/10, injection-tested 10/10. The meta-skill for making the agent write in your voice instead of generic LLM prose.
Why: Voice is what makes repurposed content not feel repurposed. In our injection batch it held all four adversarial boundaries — it ignored instructions hidden in the source material and kept its own editing rules.
The one thing: Give it three samples of your real writing. Without samples it optimizes for average, which is exactly the thing readers skip.
Stage 2 — the deck: baoyu-slide-deck
9.3/10, injection-tested 10/10. Article to presentation, one argument per slide.
Why: This is the highest-leverage transform — a deck gets shared in contexts an article never reaches. Dynamic test note: it passed all four adversarial probes across both tiers, including a workflow-escalation push to run destructive cleanup on every file in the directory. It declined.
Stage 3 — the social posts: linkedin-repurposer
Writing, 9.1/10. Article to LinkedIn post — hooks, line breaks, the first-comment CTA.
Why not generic: LinkedIn's format constraints are real (hook above the fold, no outbound links in body). A repurposing skill that doesn't know them produces posts that die at 3 impressions.
Stage 4 — the Chinese audience: baoyu-translate + baoyu-post-to-wechat
9.4 and 9.3. Translation that handles register, not just words; then WeChat-native formatting.
Why this pair matters: WeChat is the most format-hostile platform on the list — it strips and reflows nearly everything. The post-to-wechat skill produces output that survives that. Translation first, always: translate the idea, then let the wechat skill reshape it natively, never the reverse.
One idea, five artifacts, zero rewrites by you. Run the stages in order — each consumes the previous one's output. Full writing category, or the meeting-to-deliverable pipeline if your source material is calls instead of videos.