27 July 2026

How Claude Code Works, From Tokens to Agents

Detailed description of how AI coding assistants work. They’re not magic — they’re token-predicting models in loops with tools.

Some interesting takeaways:

• the model doesn’t use facts - it assigns probabilities
  • if you ask “the capital of France is…”, the most probable next token is “Paris”, but the model didn’t get that based on facts
• if you ask the model to explain how it came to an answer, it doesn’t tell you its original reasoning
  • if you need to understand why a decision was made, ask for the reasoning before the decision, not after
• bigger context doesn’t give better results
  • on the contrary, it can lead to “context rot”
  • RAG (Retrieval-Augmented Generation) helps with this - read the context and only add the relevant parts in the prompt

Part 2 repeats the concept that “context is a finite resource with diminishing returns”. The recurring theme is “context engineering” — curating what goes into the model’s window across a whole session.

Beyond Fable: Can a Local LLM Replace Cloud AI for Security Code Reviews?

Very in-depth, detailed research into how to get the best of both worlds - the power of cloud AI to design and orchestrate a security code review, and a local LLM so that you don’t have to send your code to the cloud.

If the article is too many words, watch/listen to the podcast.

Work loudly

When you work remote people don’t see the work you’re putting in - they only see the result. So you should make the work visible as you do it in places people can see it.

TimeProvider and the End of Untestable DateTime.Now

TimeProvider ends the need for custom time abstractions in .NET, giving a standard way to inject and control time in unit tests.