Weekly News Digest
An n8n pipeline that reads the week's tech news and writes the digest in my voice — or tries to.
Why I built it
Running two blogs — one professional, one personal — meant a few hours every week just reading news and writing it up. The reading didn't bother me. The writing-up did: it was the same shape of work every time, and shape-of-work is exactly what I'd rather automate than repeat.
So I built a pipeline in n8n: a weekly trigger pulls articles from NewsAPI across eight topics, Claude filters and synthesizes what actually matters, a second pass writes the post in the right voice for the right blog, and the whole thing commits straight to GitHub. No step in the middle needs me.
The part that took longest
Not the automation — the filtering. Raw news search is noisy: duplicates, press releases dressed as articles, three outlets covering the same launch. I ended up writing a prompt that prioritizes original reporting over aggregation, and asked Claude to return a short synthesis, not just a filtered list — that synthesis is what actually becomes the spine of the post.
Where it broke
It broke on voice, not on plumbing. Early drafts read like a competent stranger had written them — correct, well-structured, and not me. So in July I rewrote the whole prompt around an actual brand-voice guide instead of a generic “write a blog post” instruction: one lead story instead of a link roundup, an explicit ban on the words that make AI writing sound like AI writing, and a rule that if a first-hand anecdote would help but none exists, the pipeline leaves a visible placeholder rather than inventing one. That last rule is the one I care about most — I'd rather publish an obvious gap than a fabricated memory.
Shipping without me
The last stage fetches the current blog data file from GitHub, inserts the new post at the top of the array, and commits with real metadata — no human touches the repo between the trigger firing and the post going live.
What I'd tell someone building this
Split it into filter, synthesize, then write. Asking one prompt to do all three at once produced noticeably worse drafts than treating them as separate stages — the synthesis step in particular is doing more work than it looks like it should.