The speculative decoding trick using existing source code is genuinely clever. Most AI coding tools treat your codebase as context - Cursor treats it as a prediction shortcut.

This kind of architectural differentiation is why the 'best model wins' narrative is wrong. Tool engineering matters more than model capability for daily coding work.

I ended up building my workflow around multiple tools for this reason - Cursor for inline edits, Claude Code for autonomous execution, GPT for quick lookups. Different architectures solve different problems.

Explored the multi-tool approach: thoughts.jock.pl/p/mult…

The 90-minute RL retraining loop is the detail that stood out most. That's not just personalization - that's the tool learning your codebase in real time.