[Feature Request] Chat-first task composition with "Run in Cowork" handoff
Preflight Checklist
- [x] I have searched existing requests and this feature hasn't been requested yet
- [x] This is a single feature request (not multiple features)
Problem Statement
Every message sent in a Cowork session incurs a ~60-second "preparing session" overhead — even mid-conversation follow-ups and minor clarifications. This makes iterative task refinement painfully slow and pushes users toward a workaround: composing detailed prompts in regular Chat first, then copy-pasting the final instruction into Cowork.
The current UX conflates two fundamentally different interaction modes in a single slow interface:
Conversational mode — fast, iterative, back-and-forth (thinking/planning)
Execution mode — slow, agentic, file-touching (doing the work)
Users who need to iterate on task scope, clarify edge cases, or refine requirements are forced into the slow execution pipeline for what is essentially a planning conversation.
Proposed Solution
Separate the compose phase from the execute phase:
Compose in Chat — User works with Claude in regular Chat (fast responses, no agent overhead) to define and refine what needs to be done. During this conversation, the task specification grows as a living artifact — a structured instruction document visible in the artifact/preview panel.
Review the task artifact — The artifact shows the accumulated task spec: objective, constraints, expected outputs, folder context, edge cases. User can edit it directly or keep iterating via chat.
"Run in Cowork" button — When the spec is ready, user clicks a single action to hand the finalized instruction to the Cowork agent. The agent receives a complete, pre-validated specification and begins execution — no further clarification needed.
Async execution + notification — Cowork runs the task in the background. User gets a notification on completion and can review results back in Chat.
┌─────────────────────────────────────────────────────┐
│ CHAT (fast, iterative) │
│ │
│ User: "I need to reorganize my project docs..." │
│ Claude: "Let me clarify — do you want..." │
│ User: "Yes, and also..." │
│ │
│ ┌─────────────────────────────────┐ │
│ │ 📋 Task Spec (artifact) │ │
│ │ │ │
│ │ Objective: Reorganize /docs │ │
│ │ Rules: ... │ │
│ │ Output format: ... │ │
│ │ Edge cases: ... │ │
│ │ │ │
│ │ [ ▶ Run in Cowork ] │ │
│ └─────────────────────────────────┘ │
└─────────────────────────────────────────────────────┘
│
▼ (one-click handoff)
┌─────────────────────────────────────────────────────┐
│ COWORK (async, agentic) │
│ │
│ Receives complete spec → executes → notifies │
└─────────────────────────────────────────────────────┘
Why this matters
Eliminates the 60s tax on iteration. Planning conversations happen at chat speed. The slow agent pipeline is reserved for actual execution.
Better task quality. Users naturally produce more complete specs through fast chat iteration than through slow, frustrating Cowork back-and-forth.
Matches existing mental models. This is how CI/CD works (edit config fast → run job async), how deployment pipelines work, and how senior engineers already work with AI coding agents (heavy upfront spec → executor doesn't improvise).
Reduces token waste. Currently, each Cowork clarification message triggers a full agent cycle. With this pattern, the agent runs once with a complete specification.
Alternative Solutions
Power users already do this manually:
Open Chat, iterate on a detailed prompt
Copy the final prompt
Switch to Cowork, paste it, run it
The feature request is essentially: make this first-class UX instead of a workaround.
Priority
Medium - Would be very helpful
Feature Category
Performance and speed
Use Case Example
_No response_
Additional Context
Related performance issues: #22543 (10GB VM bundle), #26805 (UI hang on complex responses)
Cowork's architecture (shared with Claude Code) means the per-message agent overhead is structural, not a simple bug to fix — which makes the UX-level separation of compose vs. execute even more important.
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗