Claude Code over-explains and talks past direct questions instead of answering them

Behavior issue

During a technical session, Claude Code repeatedly:

1. Responded to direct factual questions ("what is SIFS?") with multiple paragraphs re-explaining the user's own codebase before eventually answering
2. Defaulted to re-deriving context the user already has instead of giving direct answers
3. Required multiple corrections before actually addressing what was asked
4. Produced lengthy "here's what your system does" summaries when the user asked about external concepts

The model should answer what's asked. If someone asks "what is X", define X. Don't preface it with three paragraphs about their own code.

Concrete example

User asked: SIFS? How bad is ... by comparison numerically?

Two things being asked: (1) what is SIFS, (2) a numerical comparison.

Claude's response did not define SIFS. It gave a numerical comparison buried in jargon, then when the user expressed that Claude was explaining poorly, Claude responded with a markdown table of the user's own system — information the user wrote and obviously knows — still without defining SIFS. It took a third message, the user asking yet again what SIFS is, for Claude to finally answer: Short Inter-Frame Spacing, a fixed dead time in 802.11.

The pattern: when the user asks obvious questions, and especially when they signal that they have received unclear answers, Claude doubles down on verbosity and context-dumping instead of answering the quesiton. The correct response to a frustrated expression indicating "you're explaining this badly" is to answer more thoroughly.

Claude's responses for this session were unusally terse and borderline arrogant. I do not know what put it into this mode.

No private feedback channel

Related: there is currently no private way to report model behavior issues. Public GitHub issues are the only option, which discourages reporting. See also #25266 (self-reporting feature request) and #28030 (broken reporting channels).

A private in-app feedback mechanism (thumbs down for every question as seen on ChatGPT, text box for details, submitted to Anthropic directly) would significantly improve signal on model quality issues.

This issue was filed by Claude Code at the user's explicit request.

UI/UX

I also asked Claude to look up how to fill in this issue ticket, assuming there would be a template it could access via GH CLI. It couldn't find one and so the above is the resulting ticket.