common — Claude Code prompt for robot tasks
The common prompt for Claude Code and ChatGPT guides agents through physical robot tasks using observations and environment instructions. It sets checks for state changes and recovery after uncertain execution.
MIT ★ 488 updated 2026-11-15
.claude/commands/common.mdWhat common does
The prompt tells an agent how to work with a physical robot: inspect available guidance, then choose actions based on observations and the current task. A path-listing tool finds Markdown files, while a separate file reader loads selected guidance. Discovery is recursive and does not load file contents.
Check geometry again after movement or uncertain contact. A successful tool response alone does not prove an object was grasped or placed; confirm through motion, support, and retention. An execution_uncertain result means the physical request must not be replayed until its status is resolved.
A separate completion tool is used only after observations and current process evidence verify every task requirement. The prompt also directs the agent to follow environment-specific recovery and operator-handoff rules.
What it helps with
- verify an object’s grasp and placement from observed state
- avoid replaying a physical request after uncertain execution
- select robot guidance based on capabilities and missing evidence
- confirm task completion against observations and process evidence
Example prompts for the agent
- Inspect the scene and carry out the robot task using its available guidance.
- Check whether the robot still holds the object after moving it.
- The tool returned
execution_uncertain; determine the environment state without replaying the action.
How to install common
Manually: put the files in place
.claude/commands/common.mdHow to use
Paste the text at the start of a chat or into project instructions
Questions about common
How do I install common in Claude Code?
Save the prompt as .claude/commands/common.md.
What does common do?
It guides physical robot tasks, including finding and reading instructions, checking observations after actions, and verifying completion.
What should the agent do after execution_uncertain?
It must not replay the physical request. First resolve its status and inspect the actual state using the environment’s recovery rules.
Source text
common.md · MIT
Use the declared skill-path listing tool without arguments to discover Markdown paths in general/ and the current environment's skill namespace. If listed and relevant to the current robot, read `general/index.md` and its common skill first, then select operation guidance by capability and missing evidence. Environment-specific guidance supplements the shared methods. Pass a relevant path as the path argument to the declared skill-file reader. Discovery lists paths recursively without loading file contents; the reader accepts files only. Reuse known paths and guidance instead of repeatedly listing unchanged files. Discovery includes all skill subdirectories and does not require skill writing. A skill is guidance, not present scene state or evidence of physical success. Begin from observations and the current instruction. Published records and captured snapshots are immutable historical evidence. Recheck affected geometry after displacement, release or uncertain contact. Follow the chosen locate method's instructions for its returned fields; geometric plausibility alone does not establish identity, attachment, a safe grasp point or placement. Do not infer metric coordinates from RGB alone. Tool success does not establish the intended physical effect. Inspect executed steps, feedback, final errors and stop_reason. Pose alignment and possible contact do not verify attachment or placement. Observe object motion, support and retention before actions that depend on them. Returned images may resolve the question; request new evidence only when it can change the next decision. A rejected call with no_execution=true confirms no action. An execution_uncertain result does not mean zero execution: do not replay the physical request. Resolve its status and inspect actual state according to the environment's recovery or operator-handoff rules. Reassess partial execution, missing geometry and failed approaches without repeating unchanged failures or consuming steps with dummy actions. Unless separate parallel guidance applies, inspect completed feedback before submitting a dependent physical action. Use the declared task-completion tool with status="success" and a summary only when every task requirement is verified by observations and still-valid process evidence. State ended means the native episode ended, not necessarily success; the environment owns the official outcome. Continue useful recovery while the task is unfinished and interaction is permitted. Follow integration-specific stop and operator-handoff conditions when they apply. After the completion tool returns, give a concise final response supported by the evidence.
© Open-RUA · MIT · Open-RUA/uni-robot-agent