[whitefables]

I was taking Claude Code (CC) and Claude Desktop (CD) apart during the weekend to understand how to solve a particular problem over the weekend on my own AI harness.

Got Claude Code to take apart the CLI (bun, Mach-O) and desktop app (Electron, Vite) app to look underneath the hood and observe how the system prompt was constructed, what tools were available, how it differs across model and how it tackle some annoying issue with the models.

Some interesting findings: - there are more than 6 different prompt variations across CC & CD - Anthropic has a section for heron_brook which allow them to inject arbitrary prompt into the live system prompt of already-installed clients - sticking issues like excessive use of bullet points and URL hallucination needs to be fixed in prompt - fast mode for Opus doesn't use a different model, just different decoding layer - CC & CD have different set of tools available entirely, CD has more computer use and browser use tools shipped with it

Have captured many different prompt snippet on how the Anthropic team deal with various issues in the full blog linked.