How I went from 15 duct-taped Make automations to a launched product
I work in competitive intelligence. My job is making sure we never get blindsided by what competitors ship.
Small team, huge market. I couldn't read everything, so I built Make.com automations. 15 scenarios pulling industry news, filtering with GPT, emailing digests to the team.
It worked. Until it didn't.
Scenarios broke silently. Only I could fix them. Every new tracking need meant another afternoon building another fragile workflow. I became the team's information bottleneck, on top of my actual job.
Then during a major industry event, all hands were on deck and the automations were sitting broken. Our CEO walked into a partnership meeting and found out our main competitive advantage had already been shipped by the exact competitor sitting across the table. Nobody caught it.
That's when I built it properly.
Summry: you pick topics, set tone and schedule, get a digest with only what's new. No scenarios to maintain. Anyone on the team sets up their own topics in 2 minutes without finding the one person who knows how Make works.
Live at summry.io. 3 topics free, paid from €5/mo. Just me and my brother, bootstrapped.
How do other founders stay on top of their industries? Curious what your setup looks like.
The GPT filtering step is usually where this breaks in subtle ways. A vague "is this relevant to topic X" prompt works until edge cases show up, and then it starts letting things through or blocking the wrong things.
One thing that helped me: breaking the filtering prompt into typed blocks (objective, constraints, examples of what to include/exclude) rather than freeform text. Each block can be tuned independently when results drift. I built flompt for exactly this, an open-source visual prompt builder: github.com/Nyrok/flompt
Yeah this is literally what killed my Make setup. Filtering prompts were opaque blobs and edge cases would slip through for weeks before I'd notice.
With Summry we pushed that complexity off users entirely - they describe topics in plain language, we own the filtering logic and test it across use cases. Our users are non-technical so giving them prompt blocks to tune would've been the wrong call. The bugs moved to our side, not gone.
Who uses flompt most in practice?