Built an automation that turns one blog post into platform-ready drafts for LinkedIn, X, and a newsletter

Built this over the weekend because I was tired of manually rewriting the same blog post three different ways every time I wanted to share it. The setup is simple on the surface, paste a blog URL in, it scrapes the article, and spits out a LinkedIn post, an X thread, and a newsletter intro, all sitting as drafts ready to review. First version used one prompt asking for all three outputs at once. Outputs were all technically fine and all sounded identical. Same structure, same voice, just different lengths. Took me a while to actually sit down and think about why, instead of just blaming the model. The reason is each platform rewards something different. LinkedIn wants a counterintuitive opener and ends on a question because that's what actually gets comments. X needs the hook in line one or nobody reads past it. A newsletter intro should read like an email from someone you trust, not a press release. Those are three different jobs, not three lengths of the same job. One prompt averaging across all three was always going to land somewhere generic. Rebuilt it with three separate prompts, each written around what that specific platform rewards instead of just length instructions. Outputs immediately stopped sounding like the same machine wrote all of them. Second decision that mattered more than I expected, whether to auto-publish once the drafts were ready. Decided against it. The output is good enough that auto-publishing was tempting, but "reads fine" and "I'd put my name on this without checking" are different bars. Built in a review step instead, drafts sit pending until I've actually read them, then I approve and they go out. Already caught a couple of off-brand lines that would've gone live otherwise. Two things I haven't fully solved building this, scraping breaks on a chunk of the internet. Paywalled articles return a login wall, JS-heavy blogs return basically nothing since nothing's rendering. Right now I'm just checking if the scraped text is long enough to be real content, which catches the obvious failures but isn't an actual fix. If anyone's solved this more reliably I'd genuinely want to know, still feels like the weakest part of the build. Also doesn't work on short posts. Anything under maybe 500 words just gives three thin versions of the same three sentences, not enough material to split into real distinct angles.