ROS 2 AI skill improves warnings more than raw accuracy
A ROS 2 knowledge pack for was tested with Claude Code, Codex, Cursor, and Gemini CLI. The same six ROS 2 tasks were run with the skill turned off and turned on, then judged blind in a separate session. The basic accuracy was almost the same in both cases.
The models already knew a lot about ROS 2. The difference was in safer first answers and better warnings. For micro-ROS on ESP32, the version without the skill produced code that could leave the device stuck after an agent disconnect, while the skill version used a reconnecting state structure.
For C++ scaffolding, the skill version added a more production-like setup, tests, and a warning about an older build method. For a QoS mismatch, both versions found the fix, but the skill version also warned that DDS can drop a connection quietly without showing an error.
Key points
- The test used a ROS 2 skill with Claude Code, Codex, Cursor, and Gemini CLI.
- Six matched tasks were tested with the skill off and on, then reviewed blind.
- Accuracy stayed roughly the same because the models already knew ROS 2 fairly well.
- The skill helped most with warnings, safer defaults, and more first code.
- A personal knowledge file may help s reduce repeated AI coding mistakes.