I’ve been a Time Team fan for as long as I can remember. There’s something deeply satisfying about watching Tony Robinson, Phil Harding and the gang rock up to a field somewhere in Somerset and spend three days finding Roman mosaics under someone’s back garden. The show ran for 20 series on Channel 4 from 1994 to 2014 and has been revived on YouTube since 2022 — that’s over 290 episodes and roughly 280 unique dig locations across 30 years.
I’d always wanted a map of every dig site. The kind of thing where you’re driving through Gloucestershire and you can check your phone and go “oh, they dug a Roman villa half a mile from here.” So I decided to build one. But rather than spending weeks doing it manually, I did it with an AI assistant — specifically Claude, running as a persistent agent called Freddy Five that has access to tools, files, and the web.
How we built it
The process was genuinely interesting. I asked Freddy to find the location of every Time Team dig over the last 30 years. It started by pulling the Wikipedia episode list (which is surprisingly comprehensive) and then cross-referencing with individual series pages, the Time Team fan database, and the Time Team Digital YouTube channel. The Wikipedia data was particularly good because many of the individual series pages included actual GPS coordinates from the dig sites.
The first pass produced 271 mapped locations with a CSV database, a KML file for Google Maps, and an interactive web-based map with Leaflet.js. Each location was colour-coded by era (Channel 4 in blue, specials in orange, YouTube revival in green) and you could search, filter, and click through to episode details. Pretty impressive for what was maybe 20 minutes of work.
The critique
Here’s where it gets interesting. I took the dataset and gave it to ChatGPT and asked it to critique the work. The feedback was thorough and honestly quite fair:
1. A geocoding error — Aston Eyre in Shropshire had been placed at imprecise coordinates. Not in Italy as the critique suggested (it was actually still in Shropshire) but not as precise as it could be.
2. Missing episodes — about 10 episodes had been missed, including some from Series 8, 12, 13, and 14 that weren’t on the main Wikipedia list page. Things like “The Leper Hospital” in Winchester, “Scotch Broch” at Applecross in the Scottish Highlands, and “The Abbey Habit” at Poulton in Cheshire.
3. Coordinate precision — most coordinates were village-level centroids rather than actual dig site locations. Fair point for a research dataset, though for the purpose of “find nearby digs while driving” it’s more than adequate.
4. No quality metadata — the dataset had no way to distinguish between a precisely known GPS coordinate and a rough village-level estimate.
5. Naming inconsistency — titles were raw from Wikipedia rather than normalised to a consistent format.
The fix
So I fed ChatGPT’s critique back to Freddy and asked it to do a v5 quality pass. It fixed the Aston Eyre coordinates, added the 11 missing episodes (bringing the total to 291 rows), normalised all naming to a consistent pattern, added coordinate accuracy grades (A for site-level, B for village-level, C for broad centroid), and regenerated everything — the CSV, KML, and interactive map.
The final dataset has 291 episodes, 281 mapped locations, coordinate accuracy grades, review flags, and a proper audit trail. One location — a 2025 Viking boat burial dig in Shetland — was deliberately left unmapped because no specific site coordinates have been published yet. I respect that kind of restraint in a dataset.
The result
Here’s the final map on Google My Maps — every Time Team dig from 1994 to 2025:
View the Time Team Dig Locations Map
Some things that jumped out once you see it all plotted:
- Somerset is the most-dug county with about 16 sites. Makes sense — it’s basically one giant archaeological site.
- London has around 14 digs, from Lambeth Palace to Vauxhall to Greenwich to Westminster Abbey.
- They went international more than I’d remembered — Spain (twice), France (twice including D-Day), the Netherlands, Belgium, Nevis in the West Indies, and three trips to the USA.
- The YouTube revival has been genuinely ambitious — Sutton Hoo, a Greek ancient city at Vlochos, and a Viking boat burial in Shetland.
What I learned about working with AI
The interesting bit isn’t really the map (though I’m quite pleased with it). It’s the workflow. The whole thing — research, data gathering, geocoding, map building, critique, quality pass — took a few hours of elapsed time and maybe 15 minutes of my actual attention. I asked a question, got a first cut, had another AI critique it, fed the critique back, and got a polished result.
The most useful part was the critique step. Having ChatGPT audit Claude’s work (and vice versa) is genuinely powerful. They catch different things. It’s not unlike having two analysts review each other’s work — a practice I’ve advocated for years at Meta. The machines are getting good enough that peer review between them produces meaningfully better outputs.
If you’re a Time Team fan, grab the map and use it next time you’re on a road trip. And if you’re interested in AI workflows, the key takeaway is: don’t treat AI outputs as final. Run them through a second model, or at least a second pass. The compounding quality improvement is real.
Now if you’ll excuse me, I need to plan a Somerset road trip.




















