A requirement stated during diagnosis is treated as an order to start editing
Summary
Claude Code treats a statement of a requirement as an order to start editing, before it has shown that it understands the problem.
What happened
After the model had measured that a deleted object came back in one configuration, the user wrote: "deleted object shouldn't come back". The model immediately began inspecting the functions it intended to modify on the user's branch, with the plan of editing them. The user stopped it: the user needed the model to understand the problem first, not to begin another round of changes. The session had already spent hours on changes built without that understanding.
Why this matters
A requirement statement in the middle of a diagnosis is usually a correction of the model's understanding, not a work order. Jumping to edits skips the step where the model states the cause and the design constraint back to the user, which is exactly the step that had been failing all session.
Expected
- When the user states a requirement during diagnosis, first state the cause and the design constraint back, with evidence, and wait.
- Start edits only when the user asks for them, or when understanding has been confirmed.
Environment
- Claude Code CLI, Opus-class model, paid subscription
- Related: #96465
---
Co-authored with Claude Code.