Decompose and Plan

The problem this solves

Local models generate code reactively — they start typing the moment they see the task. This works for trivial tasks but produces mediocre output for anything complex. Sonnet's "extended thinking" mode is essentially a forced decomposition pass before generation.

We can replicate this behavior with M2.7 by requiring it to fill out a structured plan template before any code is written. The act of filling the template is the reasoning step.

The planning prompt

You are planning the implementation of the following task. Your job is NOT
to write code yet — only to plan.