CLAUDE.md Doesn't Work (Say Researchers). But Is That True?
Why academic research on context files tells a different story than daily practice with AI coding agents.

Why academic research on context files tells a different story than daily practice with AI coding agents.


Somewhere in your organisation right now, someone is reading a 200-page contract. Not because they enjoy it. Because a clause on page 147 might conflict with a regulation that changed six weeks ago. And if they miss it, it becomes a problem that lands on someone else's desk three months from now - except by then, no one remembers why the decision was made.

You call your energy provider. Something's wrong with your bill. The agent is friendly, asks the right questions, looks things up. Ten minutes later, it's sorted.
What you don't see: that same agent then spends another eight minutes documenting the call. What was the issue. What was agreed. What follow-up is needed. All typed manually into a CRM system, while the next caller is already waiting.
I did my PhD in Nuclear Reactor Physics. Now I work with AI automation.
So when I stumbled on a video about the Siemens keynote at CES 2026, I had to pay attention.
If you've ever managed a large GitHub issue with multiple sub-tasks, you know the drill: create sub-issues one by one, set up a tracking PR, manually update status tables, ensure all the Closes #XXX statements are in sync. It's tedious, error-prone, and takes 20-30 minutes of pure overhead for each large feature.
What if you could compress that entire workflow into a single command that takes 2 minutes?