return_immediate — Claude Code prompt: what it does, how to install
The return_immediate Claude Code prompt sets how an agent handles tool calls: each result arrives before planning continues. It groups independent calls and pauses when the next decision needs fresh feedback.
MIT ★ 488 updated 2026-11-15
.claude/commands/return_immediate.mdWhat return_immediate does
The prompt has the agent process each tool result before moving on to another planning exchange. Returned state, images, telemetry, and errors feed directly into the next step. It does not use background-job handles or waiting tools.
Independent calls with known inputs can run as a batch. The agent pauses at a decision point that requires new feedback. A completed tool call is evidence that the operation ran, not proof that the physical task succeeded.
What it helps with
- process tool results before continuing the plan
- batch independent calls with known inputs
- pause when a decision needs new feedback
- separate operation completion from physical task success
Example prompts for the agent
- Follow
return_immediatewhile using tools. - Process each tool result before planning the next step.
- Batch independent calls, then pause when fresh feedback is needed.
How to install return_immediate
Manually: put the files in place
.claude/commands/return_immediate.mdHow to use
Paste the text at the start of a chat or into project instructions
Questions about return_immediate
How do I install return_immediate in Claude Code?
Place the file at .claude/commands/return_immediate.md.
What does return_immediate do?
It sets a tool-use sequence where each result is processed before planning resumes. It also distinguishes a finished operation from proof that a physical task succeeded.
Can return_immediate batch tool calls?
Yes. It allows independent calls when their inputs are known, and pauses when a decision depends on new feedback.
Source text
return_immediate.md · MIT
Every tool call returns its completed result before the next planning exchange. Consume returned state, images, telemetry and errors directly. No background job handles or waiting tools are exposed. Batch independent calls with known inputs; stop at decision boundaries requiring new feedback. Completion is execution evidence, not proof of physical task success.
© Open-RUA · MIT · Open-RUA/uni-robot-agent