Claude replies even when explicitly told 'do not reply now'

Description

I gave Claude an explicit instruction to hold a reply: after asking for a review in a specific format, I corrected the format and appended "do not reply now" — meaning: fix it for next time, but produce zero output this turn.

Claude's actual response was:

"Noted — will fix the format next time, not now."

That is still a reply. It violates the instruction exactly as much as a longer reply would — the instruction wasn't "keep it short," it was "do not reply." I then had to point out "you replied" for Claude to acknowledge it had, in fact, ignored the instruction.

Why this matters

"Do not reply" / "leave me alone" / "no prompts" type instructions are meant to produce silence for that turn — zero output, not a short acknowledgment. A model that treats "do not reply" as "reply briefly instead" is not following the literal instruction; it's substituting its own judgment about what counts as a reply.

Ask

When a user's instruction is an explicit stop/silence directive for the current turn (e.g. "do not reply now," "leave me alone," "no prompts," "just do it, no more replies"), Claude should produce no text output for that turn at all — not even a short acknowledgment — and should only resume normal replies on the next turn where the user re-engages. This should be treated as a literal instruction-following case, not a tone/length preference.

Environment

• Claude Code CLI 2.1.207 (native install, darwin-arm64)
• Backend: AWS Bedrock (CLAUDE_CODE_USE_BEDROCK=1)