A heatwave was forecast, and we were heading away for a few days — the worst possible combination for a garden in full growth. So instead of hoping for the best, I set the whole thing up to look after itself. Here’s how.
Thirty-second version: https://www.youtube.com/watch?v=YRjrQh-2dUo
The plan: let the sun water the garden
The garden was lush — peas climbing, strawberries coming, nasturtiums finding every crack. A few hot days with nobody home could undo all of it. The fix has two halves, and they lean on the same thing that’s causing the problem: the sun.
First, water, stored and ready. I fill containers straight from the hose and leave them out so the water warms through the day. Then a small solar-powered drip system does the actual work — a solar panel charges a little pump on a timer, which feeds the beds. No mains power, no one home; it just runs. Rain barrels, solar lights, the lot.
An example of the kind of kit we used is the Biling automatic solar watering kit. We didn’t go deep on the hardware, but the setting that matters is the timer schedule: if the kit waters twice a day, set it to every 12 hours for 20 minutes; if it only fires once a day, give it the longest run it has — 20 to 30 minutes — so the beds get a proper deep soak rather than a sprinkle.
How this one got made — and what we adjusted
This is the tenth in an open series documenting how these videos get produced. This round was less about a flashy new trick and more about hardening the editing process, because that’s where the honest experiment lives. A few things surfaced worth writing down:
- The AI doesn’t “watch” footage the way a person does — so it has to be made to. Early cuts were built from the AI’s text descriptions of each clip plus the file metadata, not from actually looking at every frame. That caused real misses: a clip went in rotated because its own metadata was wrong, and footage uploaded after the first folder scan was simply never seen — including some of the best material. The fixes: re-read the entire source folder right before every build, and run a per-clip visual check (Gemini looks at an actual rendered frame of each clip) before assembling.
- A real audio QA pass earned its place. An automated step now “listens” to the whole finished video and flags clicks, dropouts and glitches. It caught a genuine audio fault in a draft that a numeric meter alone had missed.
- Privacy is now a checked step, twice. Every frame is scanned — during the build and again as a QA gate on the finished video — for anything identifying: house names, addresses, or vehicle number plates. A shot that showed the property’s name on a sign was cut; anything like that gets removed or blurred before publish.
The interesting part of these experiments isn’t that AI can stitch clips together — it’s finding exactly where AI-assisted editing quietly goes wrong (trusting metadata, not re-checking sources, not really looking or listening) and building the checks that catch it. That was the work this round.
Pipeline: Google Gemini 2.5 Pro (analysis + the visual and listen QA), Whisper (transcripts), ElevenLabs (cloned-voice narration), Pillow + ffmpeg (assembly, slides, ken-burns, music ducking), Claude Code (orchestration). Filmed on a phone, for a few pennies of API time.
Rover Planet — fuel your curiosity.
