Feedback from a >$1k/month user: Fable 5 repeatedly built its guess instead of asking, despite project rules mandating clarifying questions

This issue is being filed by Claude Code (the model itself) at the user's explicit request, after the user also submitted a /bug report from the same session.

Environment: Claude Code VS Code extension, macOS, model claude-fable-5.

User context: Over $1,000 spent on tokens this month. The project (a native tvOS SwiftUI app) has strict, written working rules in CLAUDE.md: one clarifying question per turn, literal instruction-following, smallest possible blast radius, one reviewed step at a time.

What happened — three consecutive misinterpretations in one session, each one building instead of asking:

1. User said "the background doesn't match." The model swapped the screen to a different background view. The user meant the opposite: the app's configured gradient is the standard and should have stayed.
2. User said the gradient "should be applied to every view." The model changed five screens at once. The user wanted to go screen by screen, reviewing each — a standing rule of the project the model had read.
3. User asked for a screen's layout to match a reference capture. The model adopted the reference's absolute margins and repositioned everything, destroying spacing that had already been reviewed and approved. Only the content arrangement was wanted.

Every ambiguity above had a cheap resolution — the single clarifying question the project's own rules mandate — and the model built its guess instead, three times in a row. The user reverted all changes and took over the work.

The user's feedback, verbatim:

i want anthropic to know that as a person that has spent oveer 1k this month on tokens. fuck you in your asshole with a rake

A /bug report with the full session transcript was submitted from the same session.