The below are compilation of my builds + case studies ( It’s all simple stuff,nothin fancy) . Cheers !
Stack: n8n (Docker), Python (Flask), Claude Sonnet 4.5 + Haiku 4.5, Tweepy, PostgreSQL, Braintrust Status: Live and posting daily to X at 9PM IST Cost: ~$0.016 per run end-to-end Build time: ~22 hours across two sprints
For about six months I had a morning ritual that wasn't working.
Open laptop, scroll Rundown AI, then Latent Space, then Simon Willison's blog, then Hacker News, then five Twitter threads people had recommended. By 9:30 AM I had "consumed" maybe twelve stories. By Friday I had bookmarked thirty articles I never opened again. And if you'd asked me on a Sunday what actually shipped that week in AI, I genuinely couldn't tell you. I had read a lot and knew almost nothing.
The signal-to-noise problem in AI news is real. Every newsletter is racing to cover the same headline launches. Twitter is fast but exhausting. Long-form blogs are great if you have an hour you don't have. And the worst part is that even when something genuinely useful drops (a new pricing tier, a model that handles an edge case you've been working around, a benchmark that changes your stack decision), it gets buried under three hot takes about whether AGI is one year away.
I tried the existing solutions. TLDR.AI is fine but reads like a press release roundup. The Rundown is decent but optimised for scroll-time, not action. The newsletter format itself is the problem. You read it, you nod, you close the tab, and an hour later it's gone.
What I actually wanted was four tweets in the evening that told me, very specifically, what I should care about and what I should try this week. Not "AI is changing how teams work." Something closer to "Anthropic shipped prompt caching, it cuts cost ~90% on long-context calls, here is where it matters for your stack."
So one weekend I sat down to build it.
The first version came together fast. Pull from ten sources via RSS, dedupe by URL, ask Haiku to filter, ask Sonnet to write a four-tweet thread, post to X via Tweepy. Add n8n on top to fire it daily. Write a Postgres schema for dedup. Wrap the whole thing in two Flask services so the pipeline and the publisher could fail independently.
It worked. The agent started posting.
And then I read my own tweets and felt the same dull thing I felt about the newsletters. Vaguely informative. Forgettable. Lots of "this is important for builders" and "AI is rewriting the rules" energy. Nothing you'd actually do anything about.
The hard part wasn't building the agent. The hard part was that I had built a smaller, slower version of the exact problem I was trying to solve.
I changed the prompt. The tweets felt better. But "felt better" is a useless answer. By day five I had three different prompts in my notes folder and no way to tell which was actually best. I'd been a PM long enough to know that "I think this is better" is not a measurement.