Claude Cowork keeps hanging? Here's the fix that runs every session.
When Claude Cowork stalls mid-task with a decision loop, the problem isn't the task — it's a missing system prompt. Here's how to interrupt a hung session and prevent it happening again.
It happens at least once a day. I give Claude Cowork a task, it says something like "Before I build this, I need to make three decisions" — and then nothing. The spinning indicator runs. Minutes pass. No questions appear. No output. Just a frozen session that looks like it's thinking but isn't going anywhere.
If you're using Claude Cowork for anything beyond simple one-liners, you've hit this. It's not a bug you can file and wait for a patch. It's a structural behavior you need to learn to work around. Here's what I've figured out.
What's actually happening
When Claude enters that "three decisions" mode, it's not deep in thought. It's trying to call an interactive elicitation tool — a UI component that presents you with clickable options. The hang happens when that tool call stalls before it renders, usually due to context size, rendering latency, or a queue backup in the session. Claude is waiting for infrastructure that isn't responding.
The distinction matters: this isn't Claude being uncertain about your task. It's Claude getting stuck in a planning loop while the underlying tooling fails to surface. The task itself is often perfectly clear; the session just can't get out of its own way.
The hang isn't Claude thinking. It's Claude waiting for something that isn't coming.
The three fixes, in order of preference
First, interrupt it. Send a new message mid-spin. Something like: "Stop. Proceed with your best judgment, state your assumptions, and go." Agentic Claude responds to interrupts — it doesn't need to finish its current thought before reading your next message. This works more often than people expect because the model itself isn't frozen, only the tool call is.
Second, refresh and retry with a stronger opening prompt. Before you describe the task, add: "Do not ask me questions. Make reasonable assumptions, state them briefly, then build." This suppresses the elicitation reflex before it fires. You're not dumbing down your request — you're removing the ambiguity-handling pathway that leads to the hang in the first place.
Third, break the task down. If Claude is spinning on a decision loop, the task is usually too open-ended. A narrower first step — "build the data structure for this, nothing else yet" — gives Claude a concrete scope it can start on without needing to ask anything.
The real fix: your system prompt
The permanent solution isn't a retry strategy. It's your Cowork system prompt. Claude Cowork has a persistent system prompt field that runs at the start of every session. If it's blank, you're starting cold every time and relying on each individual prompt to set behavioral expectations. That's where the hangs sneak in.
I now keep something like this in my Cowork system prompt:
"Proceed autonomously on all tasks. State your assumptions inline rather than stopping to ask clarifying questions. If you reach a decision point, make a reasonable choice, note it briefly, and continue. Only pause if something is genuinely blocking — not just ambiguous."
Four sentences. It doesn't remove Claude's judgment — it redirects how that judgment gets expressed. Instead of stopping to surface a question, Claude makes a call and tells you about it inline. You can course-correct in the next message if the assumption was wrong. The session keeps moving.
This is the same principle that makes good delegation work. You don't want a team member who stops every ten minutes to check whether you meant this or that. You want someone who makes a reasonable call, flags it, and keeps going. The system prompt trains that behavior before the first task even starts.
Why this matters more than it seems
Agentic AI sessions have a compounding problem: a hang early in a session kills momentum for everything that follows. By the time you've interrupted, restarted, and re-explained context, you've spent more time recovering than the original task would have taken. The friction isn't just annoying — it breaks the flow state that makes agentic work valuable in the first place.
The system prompt fix costs you thirty seconds once. The hang costs you five minutes, every time, indefinitely. That math is easy.
If you're running Claude Cowork without a system prompt, open the settings now and add yours. Start with the four sentences above and adjust from there as you learn what your sessions actually need.
The tool is genuinely powerful when it moves. Getting it to move consistently is mostly a configuration problem, not a capability one.
Sources & further reading
External
- Anthropic — How the agent loop works (Claude Code Docs)
- Ralphable — The Claude Code 'Infinite Loop' Bug: How to Spot It, Stop It, and Get Your Project Back on Track
- Claude Code Autonomous Mode: Guide to permission modes and autonomous operation
Related posts