The worked atlas¶
A transport atlas of Leeds city centre, planned and built by Claude, an AI assistant, and published as this website.
Why this section reads differently
Everything else on this site is written to be plain. Short sentences. Simple words. Nothing decorative. That is a deliberate choice for instructions you have to follow while also learning to program.
This section is not written that way, and the contrast is the point.
It is what a capable assistant produces when asked to plan a large piece of work and then carry it out at length. It is longer, more confident and considerably more pleased with itself than the rest of the course.
Read it for two things. First, what you can build in five weeks with an assistant and the habits weeks 1 to 3 teach. Second, the difference between writing that sounds rigorous and work that is rigorous — and how you would tell them apart.
The rules it was given¶
The same ones the project brief gives you.
- Write down what the assistant will get wrong, before delegating.
- Ask for the plan before the code.
- Count the rows before and after every filter and every join.
- Open every figure and look at it.
- A hand-check anchored outside the data is worth more than twenty that are not.
- One command rebuilds everything from nothing.
It wrote the plan first, including a list of things it expected to get wrong about itself and eight numbered predictions it might fail. Then it executed the plan. Where the plan turned out to be wrong, the chapters say so rather than quietly correcting it.
What it built¶
-
Written before any code existed. The invention list, the architecture, the hand-check register, the risks, and eight falsifiable predictions.
-
Chapter 1 — The patch and its stops
Where you can catch something, and where you cannot. Contains the most instructive mistake in the build.
-
Two years of STATS19. Four integers decide whether the whole chapter is true.
-
Where the patch sits in England's range, and why decile 1 is the poor end.
-
The chapter the rest of the atlas is about — and the join it refuses to make.
-
The only chapter whose subject does not exist yet.
-
What the network reaches. Three ways to be refused by an API.
-
8,760 hours, returned as parallel arrays with nothing holding them together but position.
-
All seven chapters together, every figure on one page, and what they say when read as one document.
-
27 hand-checks. 14 anchored outside the data. Generated from the register, so the honest number cannot be inflated.
-
Two lines to change. How the website is generated, and how to publish yours.
Four things it got wrong¶
Kept in, because they are more useful than the parts that worked.
A query that returned a confident, incomplete answer
Chapter 1 fetched ATCO area 450 and got 1,314 bus stops. The map looked complete. Rail stations are in area 910, a separate national pseudo-area, so the atlas of Leeds public transport had no Leeds Rail Station in it.
Nothing failed. No traceback, no empty table, no zero count. The row counts could never have caught it, because the missing rows were never there to count. A hand-check written before the code — a station I know exists must be in this file — caught it.
Ten stops drawn nowhere
The first stop map had two categories, bus and rail. Ten access points are
typed RSE, a station entrance, and belonged to neither. They were
fetched, filtered, counted — and drawn on no figure.
The only visible trace was a legend that added up to ten short of the total.
A figure that ran perfectly and showed nothing
Chapter 6 asked what share of everyday destinations sits within 400 m of a stop. The answer is 100%, correctly: the furthest school in the patch is 366 m from a stop.
So the first figure was six bars, all at 100%, neatly arranged and completely uninformative. A threshold everything clears is not a measurement. The figure was replaced with one that shows where the categories actually differ.
A source that was down on the first try
The deprivation file link on the current gov.uk page returns 404. Without checking, the chapter would have fallen back to a cached copy and reported months-old numbers while appearing to be live.
That is why every page states live or cached copy for every dataset, rather than just naming the source.
The honest number¶
The agent guide records what happened when an assistant built an atlas like this before: twenty hand-checks, a banner reading "20 of 20 pass", and about four that were worth anything.
This build registers 27 hand-checks, of which 14 are anchored outside the data — against a map, a published statistic, a physical fact.
The scorecard is generated from the register rather than written, so that number cannot be quietly improved.