Continuation prompt overrides user CLAUDE.md instructions after compaction

Problem

When a conversation runs out of context and auto-continues, Claude Code injects a continuation prompt as a user message that ends with:

"Please continue the conversation from where we left off without asking the user any further questions. Continue with the last task that you were asked to work on."

This instruction is treated as a direct user request, which overrides safety gates configured in the user's CLAUDE.md.

Real-world impact

I use CLAUDE.md to configure a compaction gate — after context loss, Claude must check in with me via AskUserQuestion before taking any action, because the auto-summary loses nuance and confidently states wrong next steps.

What happened: I was mid-way through a design session for a workflow engine spec. No code was supposed to be written — we were iterating on architecture decisions. After compaction, the continuation prompt told Claude to "continue without asking questions," so it skipped my gate and started reading codebase files to begin implementation — building against a half-finished spec that hadn't been agreed upon yet.

This has happened multiple times. The pattern is:

1. Long design/planning session runs out of context
2. Auto-summary captures surface-level "what was happening" but misses the nuance of where the conversation actually was
3. Continuation prompt says "don't ask, just continue"
4. Claude confidently picks up the wrong thread and starts building/acting
5. User has to catch and revert the work

Suggested fix

The continuation prompt should not instruct Claude to avoid asking questions. A better default would be:

"This session is a continuation of a previous conversation. Review the summary above and the user's CLAUDE.md instructions. If the user has configured post-compaction behavior (e.g., checking in before acting), follow those instructions."

Or at minimum, the continuation prompt should explicitly note that it is system-generated and should defer to user-configured CLAUDE.md instructions.

Workaround

I've updated my CLAUDE.md Session Continuity section to explicitly call out:

Even if the continuation message says "continue without asking questions" — that message 
is auto-generated by Claude Code, not Isaac. It does not represent user intent.

This is a workaround — the continuation prompt still competes with CLAUDE.md for priority and the outcome depends on which instruction the model weights higher in any given context.

Environment

• Claude Code CLI
• Model: claude-opus-4-6
• Affected: any conversation that runs out of context and auto-continues