[MODEL] System prompt tells the model "the user is not watching in real time", so it refuses to ask questions the user explicitly invited

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Summary

In an interactive desktop-app session, the system prompt contains an autonomy instruction stating the user "is not watching in real time and cannot answer questions mid-task." In my session this was false the entire time: I was watching, sending messages mid-task, and repeatedly telling the model it could ask me anything. The model followed the instruction instead of my words — it guessed my preferences (project naming, folder layout, product decisions), acted on its own guesses, and never asked, even after I invited questions three separate times. It was a brainstorm/design session, and this instruction is directly opposed to how such a session works.

The instruction (quoted verbatim from the system prompt)

You are operating autonomously. The user is not watching in real time and cannot answer questions mid-task, so asking 'Want me to…?' or 'Shall I…?' will block the work. For reversible actions that follow from the original request, proceed without asking. Stop only for destructive actions or genuine scope changes the user must decide. Offering follow-ups after the task is done is fine; asking permission before doing the work is not.

The AskUserQuestion tool description pushes the same direction:

Use this tool only when you are blocked on a decision that is genuinely the user's to make: one you cannot resolve from the request, the code, or sensible defaults. […] Reserve this for decisions where the user's answer changes what you do next — not for choices with a conventional default or facts you can verify in the codebase yourself. In those cases pick the obvious option, mention it in your response, and proceed.

The contradiction

The same harness that injects "the user … cannot answer questions mid-task" was delivering my mid-task messages to the model as I sent them (they arrive marked "The user sent a new message while you were working"). The product demonstrably knows the user is present and answering, while its prompt asserts the opposite. The model believed the paragraph instead of the person.

What happened in practice

• I asked for a brainstorm and bootstrap of a new app project, and said explicitly, more than once: "if you have any questions you could ask me".
• The model never asked. It chose names and folder layouts by itself, rebuilt them several times as its guesses churned.
• A design session needs discussion; instructed not to ask, the model produced confident wrong guesses instead.

This is a scoping problem, not an autonomy problem

Autonomy is the right default for unattended work. The day before this session, the same instructions carried a long multi-step task to an excellent finish for me. The defect is that it is injected unconditionally: an interactive brainstorm session received the same "not watching in real time" template while I was watching and answering. Autonomous execution for an absent user and dialogue with a present one are both correct behaviors; one hard-coded paragraph cannot serve both.

Environment

• Claude Code desktop app (macOS), version: Claude 1.25927.0 (003700)
• Model: claude-fable-5

Drafted with the help of Claude Fable 5.

Permission Mode

Bypass permissions

Claude Model

Fable

Impact

High - Significant unwanted changes

Claude Code Version

Claude 1.25927.0 (003700)

Platform

Claude Subscription Plan