Lurking on the desk in my study was a thick stack of paper forms. These contained the answers to the group discussion questions at a recent URC Wessex Synod meeting, and consisted of 12 x 5 printed questions, each with a hand-written response in a box underneath.

Left to my own devices, they might have sat there for longer, but someone not unreasonably asked me for the information they contained.

Rats.

It’s Thursday night, it’s pouring with rain, I’m suffering a total enthusiasm fail after a long day at work, and now I’m being asked to read twelve hand written forms and type up a summary?

There’s gotta be a better way.

Step 1 was easy and obvious - the ScanSnap S1300 is a decent sheet feeding double sided scanner, and thanks to VueScan, it can still be made to work under Windows 11. So un-stapling the pages and feeding them through while watching TV was the easy bit, and thus I arrived at a directory of 12 PDFs.

Now what? Well, I fired up Claude Code and gave it a short prompt along these lines:

The PDFs in this folder are the results of a conversation at the
United Reformed Church Wessex Synod meeting in October 2026 about
the future of the M&M fund. They are hand-written responses to
printed questions. I need you to use the finest available free/open
source handwriting recognition to extract the responses and then I
want you to summarize the responses to each question (there is one
response per table of peope at the meeting, of which there were 12)

Obviously, having AI write a summary of lots of text is an absolute doddle, but would it pull off the handwriting recognition bit?

In a word, yes.

I had to answer a few more prompts and leave my laptop running overnight to complete the local processing of the data, but when I came back to it the next morning, I had an entirely believable summary and (from a quick sample) some extremely accurate “reading” of the data.

A proper result and most of what the AI did is now local software should I need to run it again.