AI found the campground I couldn't name and measured the only site left

I could remember the campground in ridiculous detail except for its name.

Somewhere over the Colorado line in Utah. Maybe Thompson City. A little blue hotel. Tent camping. RV sites. Old buses people could sleep in.

That peculiar description was enough. AI found Desert Moon in Thompson Springs and showed me the blue building and vintage bus. I recognized it and sent the website to Jason.

We needed six nights, October 4 through October 10, so I compared Desert Moon with RV Ranch in Clifton and Canyon View in Grand Junction. Desert Moon was $210. RV Ranch was $390 for the 50-amp service we needed. Canyon View started around $414 and generally wanted RVs 15 years old or newer, with possible photo approval for older rigs.

Our 1998 Holiday Rambler doesn't look like a disaster. It also doesn't look like the newer motorhome they seemed to have in mind. I crossed Canyon View off rather than make the Rambler audition for a campsite.

The remaining comparison went to Jason. He leaned toward RV Ranch but was fine with either. I leaned toward Desert Moon. We talked in person and chose the weird little place in Utah.

Only Site 5 was open. Before booking it, I wanted to know whether a 32-foot motorhome and the new dog fence would fit without putting us on top of the neighbors.

I provided the campground drawing and a Google satellite image with a 50-foot scale bar. The first result looked polished until I compared it with the real image. The AI had reconstructed the background, invented tidy campsite boxes, and drawn boundary lines directly through RVs already parked in the photograph.

“You did that very very very poorly,” I told it.

That may be my most accurate prompt of the week.

The assistant stopped and explained what had gone wrong. It had treated the job as image generation when I needed measurement and annotation. A plausible-looking campsite was useless. The Google screenshot needed to remain exactly where Google put it.

I asked whether the job belonged in Codex. It did.

The handoff carried strict instructions: preserve the image pixel for pixel, map Sites 1 through 9 before touching Site 5, use the scale bar as the ruler, separate measured facts from inferred boundaries, draw the 32-by-8.5-foot Holiday Rambler to scale, and only then test the 16-panel fence.

Codex measured the scale bar at 94 pixels for 50 feet—1.88 pixels per foot. A site-specific reservation listing described Site 5 as 29 by 50 feet. The best fence shape used eight panels along the outside and four at each end, with the RV serving as the fourth wall. That produced a yard about 18.7 by 9.3 feet, or 174 square feet.

The plan worked if the RV shifted about three feet away from the fence side. The estimate left 7.3 feet to the inferred Site 6 boundary and 3.9 feet from the outside fence rail to the inferred Site 4 boundary.

There was still a caution. Desert Moon's generic page advertised 20-by-60-foot sites, while the site-specific listing said 29 by 50. Those boundary lines don't exist in the satellite image, and no overlay can turn an inference into a survey.

We booked Site 5. The email concierge caught the confirmation and put it in Temporary Holding until after the trip.