Testing Claude Code Learning mode: custom output style leaves TODO(human)

On November 12, 2026, an author published hands-on testing of the Learning output style in Claude Code 2.1.299. The experiment was designed to check whether the style would cause the assistant to record follow-up tasks in a structured form rather than silently moving on.

In the first part of the test, the author ran the command line interface through claude -p four times with the Learning style in place. Across all four runs, the tool did not produce a TODO(human) entry, meaning the pending items were not captured in that format. To see whether the behavior could be shaped, the author then wrote a custom style file of 20 lines and applied it to the same workflow.

With the custom style file loaded, the author repeated the runs three times, and each of the three runs did leave a TODO entry in place. The write-up also documents how the style file is configured and the handoff check process that follows once the output is returned. At this point, the testing record and the setup steps are presented as a practical walkthrough for others who want to reproduce the same configuration in Claude Code 2.1.299.