Had a bit of fun with CoPilot

[mendip_discovery]

At work we had a list of items a customer wants calibrated in a scanned PDF file. I wanted to import this to Excel.

Its a scanned image of a table. There are some hand written notes and some highlighter. But mostly intact.

Tried Excel's own import tool and it failed. So I thought I would try Copilot.

It said it needed to ask a few questions, it seems to have got the headings. It then said give me a moment and it will be done.

So I gave it 5 mins and nothing.

I asked how it was getting on, it them asked some more questions then said it will be just a moment.

This cycle happened 5 times. Then I left it for a good 2 hours. There was even a claim of 'i am not stalling'.

Then finally it said it couldn't do it because its a copyrighted list. But it would offer to help me write some python code to extract the data.

So I gave up. Its not too much of an ask. If it was a little more upfront about not being able to do it then I would have been ok.

[paulca]

You spent 2 hours showing a blindman a photo and wondered why he couldn't draw you a copy of it.

Either take it up on the Python offer, ask it about OCR conversion tools, even Google Drive and meet it closer to halfway.  Also direct to excel is asking for trouble as it's proprietary.  Try a more open intermediary first.

[kripton2035]

I've given Claude Code a scan of some handwritten notes I'had taken at a meeting one year ago and that I did not recall all of it...
it read it and understand it so well it could explain me the things I did not remember about it !

[mendip_discovery]

3 column table, scanned in. OCR would do it easily, but I don't have OCR software on the work laptop. I tried Excel's Import Data from a PDF, and that normally works, but this time it didn't. So I tried AI as it's meant to be able to do lots of clever stuff, so I keep getting told.

I would share a sample, but it's got names, etc., on it. I did take a screenshot of the file, and the snipping tool offered OCR, it worked great. So there is more to Copilot being vague and avoiding stuff. Will see what its like in a few generations.

At least I didn't let it loose on my software and let it delete all the data and the backups as it decided to fix stuff.
https://www.theregister.com/2026/04/27/cursoropus_agent_snuffs_out_pocketos/

[coromonadalix]

claude use tesseract ocr, it can be installed and be used outside claude ...

[5U4GB]

Either take it up on the Python offer,

Friend of mine, a Claude fan, asked it to write him some code to convert a file from format1 to format2.  It came up with (approximately):


file1 = fopen("file.fmt1", "r");
file2 = fopen("file.fmt2", "w");
while(!feof(file1))
  fputc(file2, fgetc(file1));
fclose etc;

[paulca]

Either take it up on the Python offer,

Friend of mine, a Claude fan, asked it to write him some code to convert a file from format1 to format2.  It came up with (approximately):


file1 = fopen("file.fmt1", "r");
file2 = fopen("file.fmt2", "w");
while(!feof(file1))
  fputc(file2, fgetc(file1));
fclose etc;

If you didn't specify the formats save for the file extension and asked it specifically for code, that is a a valid file copy method with rename pattern (just with no error handling and some buffering would be nice).

The risk is actually that a human engineer is likely to frown at me, smile bemusedly and ask, "Why?"  The AI might not.

Also.  What if the next prompt is. "Great, now can we add a "filter_by_offset" method"