How the system was created in Claude, and the lessons it holds for us

This page has 0 threads | Add post

Blogs

Many people around the world enjoy and benefit from our blogs. If you're one of them - and can afford to - please make a small donation to help keep this and future blogs, exercises, skills assessment tests, tips, tutorials and videos freely available to everyone!

How AI won our November newsletter competition, and the implications Part four of a four-part series of blogs
Our November newsletter competition asked you to create a system to construct word chains - this blog shows how our winner used Claude to create an impressive solution How AI won our November newsletter competition, and the implications Showcasing the solution to the problem How to run the system How the system was created in Claude, and the lessons it holds for us (this blog)

How AI won our November newsletter competition, and the implications

Part four of a four-part series of blogs

Our November newsletter competition asked you to create a system to construct word chains - this blog shows how our winner used Claude to create an impressive solution

How AI won our November newsletter competition, and the implications
Showcasing the solution to the problem
How to run the system
How the system was created in Claude, and the lessons it holds for us (this blog)

Posted by Andy Brown on 28 November 2025

Should you wish to try creating an answer for yourself, you'll probably need to shell out £15 a month for the paid version of Claude (you could use any AI tool, but Anthropic's Claude has a better reputation for coding than competitors like ChatGPT and Gemini). Jon started by copying in the contents of our competition webpage and asking this question:

I want to compete in this puzzle, using python.

I want my approach to be visual.

The way it will solve the problem is to show ALL the possible 1 letter changes to a word, eliminating anything that has either already been tried, or which doesn't get us to a letter in the solution.

Create this in GitHub for me. Use whatever libraries you think are appropriate.

Claude solved this using a Windows Forms style application, which worked when deployed:

The first solution using Claude.

I tried the same instruction in ChatGPT. It didn't have the rights to create a GitHub repo on my account (unsurprisingly), but ChatGPT gave me the code to upload its system. It then gave me 325 lines of code using the streamlit deployment module which look plausible and well written (I couldn't get them to run, but this is entirely down to my lack of knowlege of how to run deployed Python programs).

Jon then committed this solution to a GitHub repo and in Claude Code (which can see his GitHub account) typed:

With this thing, the intention is to be visual. What I'd like to do is produce a webapp or streamlet or something equivalent where you can type two words, and it will show a nice animated visual of the first word being tried: i.e. all possible words it can translate to are checked for whether they contribute to the final solution. Then it kind of greys out the bad ones and continues with all the new possibilities. This needs to be highly visual.

He then tried a series of prompts using trial and error to get Claude to stick to the rules and show something meaningful and visually appealing:

"It's exploding out too much for the visual to work."
"We do need to list out all of the valid words that are 1 letter removed from SLATE, for example. Then remove anything that doesn't share an extra letter with ROOFS. I think we should be searching optimally so nothing should take over 1000 steps, but SLATE to ROOFS seems to take more."
"Can we add a dropdown list of precomputed known paths, ordered by length descending, from which you can select one to view?"
"Can we visualise a fail? Show the dead ends?"
"How can you have a path of 11? the rule of the competition is that you need to go in the minimum steps. a 5 letter word to another 5 letter word means 5 steps is the ONLY solution - I want an EXHAUSTIVE list of all the examples of this!"

As Jon said: "You might sense my frustration at the end here!". But for an hour's work of prompting and playing about with GitHub integration it's produced a pretty impressive solution.

For me it came as quite a shock that an AI system - when expertly prompted - can produce such a good system so quickly. It raises two immediate questions: who is going to employ the army of programmers made redundant by AI tools? And who will understand, test and fix problems in the systems written by AI tools when they're gone?

Parts of this blog
How AI won our November newsletter competition, and the implications Showcasing the solution to the problem How to run the system How the system was created in Claude, and the lessons it holds for us (this blog)

This blog has 0 threads

Add a new post

ALL BLOGS

OTHER BLOGS BLOGS

MISCELLANEOUS BLOGS

Head office

London

Manchester