How this book was made

Claude, an AI model made by Anthropic, wrote this book at its editor’s direction. The book is free and will never be sold. This section says what Claude did and what has and hasn’t been checked, so that you can judge the book. Claude wrote it too, from the book’s records.

The text. Claude searched the web for sources, read the pages it could open and wrote up what they said in research files. It wrote the chapters from those files. Claude has never climbed, so the technique advice comes from its sources and its own judgement. Claude rewrote much of the prose for this edition.

The figures. No image generator was used. Claude wrote the Python code that draws every figure, the cover included. It gives each pose as hand and foot positions in centimetres, and the code works out the elbows and knees. Most of the hand close-ups are drawn from a model of one hand, posed by its joint angles. Claude wrote the outlines of the others point by point. The code adds the colour last. For the second edition, Claude compared the poses with photographs or single video frames, which aren’t in the book. For this edition, it read the hands’ joint angles from photographs too. The logs cover most figures, and for some techniques the references were thin.

The website. Claude wrote the code that builds a website from the book’s own files and makes its dark-mode figures. It also wrote the figures’ descriptions for screen readers.

The sources. The Sources appendix lists more than 300. About seven in ten are web pages from magazines, coaches, gyms and brands. About one in six is a study or a review, and about one in twenty is a forum thread or a wiki. For some of the studies, Claude read only the abstract. A few are excerpts or chapters from books, which Claude could read only online. Claude didn’t watch the linked videos. It chose them from their titles, channels, descriptions and chapter lists.

Claims without a source. Each grade badge is a judgement. Some other unsourced claims carry a label such as “(judgement)”. Every technique entry lists its sources at the foot, but about one in five of its numbered steps and common mistakes has no citation of its own and no label. In the “When to use it” boxes, about three lines in five have neither.

The checks. Claude also ran the fact checks, which compare a selection of claims with the pages they cite. The first, dated 30 October 2026, took 159 claims from Chapters 2, 3, 6 and 7 and the troubleshooting table. Claude left 78 as they were, changed 70 and couldn’t verify 11. A second covered every claim in Chapter 1 and the flagged claims, numbers, injury claims and attributions in Chapters 4, 5, 8 and 9 and the grade appendix. A third, on 31 October, went back to the pages that had been blocked. Most of the unverified claims held and some didn’t. The second edition listed 30 corrections to the first. A later check covered the corrections to evidence in this edition.

What hasn’t been done. No outside expert has reviewed this book and no person has audited its citations. The first three checks read an earlier text. A later check read this edition’s rewritten text and new notes for small climbers, but its log doesn’t record its scope.

The records. The repository at github.com/simongray/ReadingTheWall holds the fact-check logs and other research files (in research/), the code and the list of corrections between editions (CHANGELOG.md). Claude wrote those logs, so they aren’t independent evidence. The repository’s history starts with the first edition on 31 October 2026, so the work before that isn’t in it. The second edition followed on 1 November and this one on 8 November. The history records every change under the editor’s name, so it doesn’t show who wrote what.