Two years ago I wrote a short post on Rover Planet asking whether UK hedgerows actually sequester carbon. The conclusion at the time was honest: the obvious answer felt right but I couldn’t pin a real number to it. The piece ended with “the quest for a reliable and confirmed calculator is in progress.” This week’s video is what I’d have wanted to read then.
A vertical Short version of the same material is up at https://www.youtube.com/watch?v=8ll2V7wJFOc for anyone who’d rather watch in forty seconds.
The numbers themselves
Three things stood out from the data once I went looking properly:
- UK hedgerows currently store roughly 9 million tonnes of carbon between them — carbon already drawn out of the atmosphere and held in standing biomass.
- A single kilometre of mature hedge actively sequesters about half a tonne of CO2 every year, locked into stems, leaves, roots, and the soil beneath the hedge.
- The Climate Change Committee’s recommendation is a 40% expansion in UK hedgerow length by 2050, which would push annual hedgerow sequestration to roughly 1.5 million tonnes of CO2 per year.
And underneath all of that, a single well-managed kilometre of hedge supports over 600 plant species, 1,500 insect species, and 65 bird species. Carbon storage is one job. Habitat is the other. They come together for free.
What changed in the production pipeline this week
The previous video in this series — the Carbon Impact piece on preventing emissions versus sequestering them — shipped with auto-captions, a custom thumbnail, and a Final Cut Pro project file alongside the MP4. This week tested four new capabilities against that baseline:
1. Voice realism. The narration on the previous videos used eleven_turbo_v2_5, ElevenLabs’ fast-but-not-quite-flagship model. For this video I switched to eleven_multilingual_v2 with stability dropped to 0.3 and similarity_boost raised to 0.85 — settings tuned to favour natural variation over consistency. Both versions were rendered. The tuned version is what’s in the final cut. It’s slightly slower-paced and more clearly inflected; whether it sounds more like me is for someone else to judge.
2. AI B-roll via fal.ai. Three Kling 3.0 clips were generated through fal.ai’s API: a slow aerial through a forest canopy, hands placing scraps in a compost bin, and a close-up of a sapling being planted. Two of the three ended up in the final cut, sandwiched between still-image scenes. Around two dollars of fal.ai credit for the set. The result is good enough that the question for next time is no longer “will the AI clips look acceptable” but “at what density of AI clips do we get the most signal for the lowest cost.”
3. An animated count-up. One stat card animates the “1.5 Mt CO2 per year” number from zero up to its final value over about two and a half seconds, with an ease-out curve. Rather than set up the full Remotion (React/TypeScript) pipeline, this version rendered a frame-by-frame Pillow sequence and let ffmpeg sequence them. The Remotion spike will come on the next video where the visual sophistication actually warrants the heavier setup.
4. Music. Almost. The plan was to add a music bed and duck it under the narration. ElevenLabs’ Music API was the first attempt — my current account tier doesn’t have the music_generation permission, so that route is gated behind an upgrade. The fallback was fal.ai’s musicgen endpoint, which accepted the job but was still queued at publish time. Music is deferred to next week, this time without a third false start.
The cost and time
- API spend on this video: roughly £1.50 — two ElevenLabs narrations (one baseline, one tuned), plus three fal.ai Kling clips at about £0.40 each.
- Time from blank page to two videos live on YouTube: about four hours, including the music detours.
- Runtime: 1:15 landscape, 0:42 Short.
Three skills now live
One thing that didn’t make it into the previous post: the underlying scripts are now packaged as user-level Claude Code skills so they’re invokable from any project. /rover-video builds a brand-aware landscape + Short pair, /rover-youtube uploads to any registered channel with per-channel OAuth, and /rover-wp-publish publishes the companion post to roverplanet.co.uk. The point of extracting them was to keep this pipeline from being trapped inside one project — the Flix and Kitchen Sink properties on the Rover side could plausibly use the same tooling with their own brand JSON.
What’s queued for next time
In order of priority:
- Music bed, for real this time — either via an upgraded ElevenLabs tier with
music_generationenabled, or by waiting out the fal.ai queue properly. The hypothesis is that this is the single biggest perceived-quality jump still left on the list. - Remotion (proper) for the count-up animations, on a video where the stat density justifies the setup.
- A Veo 3.1 hero shot — one expensive cinematic opener as a quality-ceiling reference.
- Voice realism iteration two — if the current tuned settings still don’t feel right, the next step is re-cloning with better source recordings rather than upgrading the model further.
The bigger thread
The thing that keeps surprising me about this exercise is the rate at which the cost of shipping decent video is collapsing. Two videos this week, a few quid in API credit, a couple of API hiccups along the way, and the result is in a place a working person on a real channel could be proud of. The constraint has stopped being budget or production capacity. It’s now sitting somewhere quieter — in the writing, in the choice of what’s worth making at all, and in the patience to keep the substance honest. Which, for a channel called “Fuel Your Curiosity,” is probably the constraint that should have always been the limiting factor anyway.
Next video drops with music and another data-led topic from the Rover Planet backlog. If anything in here is useful to your own work, the pipeline scripts are documented and the cost numbers should be reproducible. The substance — that’s still on each of us.
