Player not loading? Watch on YouTube
This Japanese-language review compares context-mode and RTK through repeated Claude Code tests. The speaker compares context-mode with RTK as tools for reducing the output a coding assistant receives. context-mode runs code in a separate sandbox and returns its standard output to the conversation. It can store larger outputs for later search through SQLite FTS5, though the speaker warns that storage depends on output size and how the tool is called. RTK instead compresses command output before passing it to the model.
The review reports 90 Claude Code runs using the same tasks and models across conditions. Its measurement compares tool-result bytes entering the conversation with an unmodified baseline, so the results are not directly comparable with the README's advertised 98% reduction. Claude Code already used scripts to aggregate data in several tasks. One Opus browser task showed roughly 90% savings, but RTK produced the same ratio. Other results ranged from reductions to nearly threefold increases. Accuracy showed no substantial difference in these tests, which used three runs per case and did not test Cursor.
The speaker also examines hook coverage and reports that external MCP responses can still enter the conversation directly. Project-scoped installation reportedly wrote user-level registration and hooks, contaminating initial comparison runs. At the September 18 recording, the presenter reported npm version 1.0.169 with its latest publication dated June 29, 2026; this is the review’s repository snapshot. Its practical advice is to request aggregate results through scripts, preserve source text when needed, and measure whether smaller outputs still answer the task.
