Pencil sketch of a voice recorder, cycling gloves and map on a cafe table with a laptop in the background

Later this year I’m cycling seven hundred kilometres across Burundi over the course of a week. I expect the trip to be extraordinary — the kind of experience where you see and feel more than you can possibly hold in your head, and by the time you get home half of it has already slipped away. I wanted a way to capture it as it happened, and turn the raw experience into something lasting.

So this was the test run. The Bright Stream started as a handful of voice memos recorded on a small dictaphone while cycling through the Cotswolds. It ended up as a five-thousand-word illustrated travelogue. The entire process — from raw audio to published page — was a collaboration between me and AI. If it works for forty miles of Oxfordshire, it should work for seven hundred kilometres of East Africa. Here’s how it went.

The ride

I carry a small voice recorder in my jersey pocket. I talk into it — observations about the landscape, things I notice, idle thoughts. It’s not a polished narration. It’s breathless commentary on hills, half-finished sentences interrupted by potholes, and the occasional “oh, that’s gorgeous” when a view opens up. The kind of thing you’d never publish as-is.

This particular ride was a forty-one-mile loop from Witney through the Windrush Valley to Northleach and back. Three hours. I recorded about twenty minutes of fragments across three files — two short clips at the start and a longer running commentary for the main body of the ride.

Transcription

I have a transcription pipeline that runs on AWS — upload the audio files, and it runs them through speech-to-text with speaker identification, then stitches multi-file sessions together with continuous timestamps. The raw transcript that comes back is usable but messy. Wind noise garbles words, place names get mangled (Clanfield becomes Crownfield, Filkins becomes Philins), and the speaker identification doesn’t always know what to do with a single voice on a windy road.

From transcript to travelogue

This is where Claude came in. I gave it the raw transcript and asked it to turn my voice notes into a proper piece of writing — something in the style of BBC Radio 4’s Open Country or the long-form travelogue tradition of Robert Macfarlane and Roger Deakin. The brief was specific: keep my voice, but make it prose. Don’t lose the personality. Where I say something interesting but half-formed, develop it. Where I get things wrong, correct me gently.

Claude did three things simultaneously:

  1. Researched the style. BBC Radio 4 travelogue conventions, the narrative techniques of nature writing, how present-tense first-person prose handles the balance between personal observation and factual content.
  2. Researched the route. Every place name, every church, every long-distance path I mentioned. The history of Northleach as a medieval wool town. The reintroduction of red kites. The geology of the Cotswolds (I’d guessed glaciers on the ride — I was wrong, and the real answer is more interesting). The Beeching cuts. The ecology of grey herons. All double-sourced where possible.
  3. Wrote the piece. Six sections following the arc of the ride, weaving my observations with the researched material. My speculation about glaciers is preserved in my own words, then gently corrected. My offhand mention of Turkdean and Indiana Jones gets expanded into a properly researched aside. The structure mirrors the ride: effort and exhilaration in the middle, reflection towards the end, a brief coda where I decide not to stop for breakfast.

The research turned up things I wouldn’t have found on my own. I didn’t know that Scireburne means bright stream, or that Colonel d’Arcy Dalton spent fifty years preserving Oxfordshire footpaths, or that a single oak tree supports over two thousand three hundred species. These details are what lift a cycling blog post into something richer — and they came from the AI doing the kind of deep, parallel research that would have taken me days.

Illustrations

Claude wrote detailed prompts for pencil sketch illustrations to accompany each section — specifying composition, subject, and a consistent graphite-on-white-paper style. I generated the images using Google’s image generation. The consistency of the sketch aesthetic across all seven images was important: they needed to feel like they belonged together, like a series of drawings from the same sketchbook.

The page

The HTML and CSS were built to feel like a premium editorial piece — warm cream background to match the sketch paper, serif typography for long-form reading, images breaking out wider than the text column. Claude built the page, I reviewed it. The whole site was deployed to AWS in the same session.

What’s mine and what’s not

The cycling is mine. The observations are mine. The voice — the curiosity, the wrong guesses, the “there will be no walking today” — that’s mine. The prose, the research, the structure, the illustrations, and the web page are Claude’s (and Google’s, for the images). I think of it as a collaboration: I provided the raw experience and the editorial direction, the AI did the craft work of turning it into something polished and publishable.

The whole process — from uploading the audio files to having a live website — took a single afternoon. That still surprises me.

What’s next

Burundi. Seven hundred kilometres, a week on the bike, a landscape and a culture I’ve never experienced. If twenty minutes of voice notes from the Cotswolds produced five thousand words, I can’t imagine what a week of riding through the hills of East Africa will generate. The voice recorder is coming with me. The pipeline is ready. We’ll see what comes back.