Ralph Krysler 🪓

I ran a little experiment yesterday to compare the difference between one-shot prompting and task-based prompting I needed to refactor the first version of the @TaskmasterAi MCP because the tools were returning the whole CLI output instead of just the raw data (which would waste users’ tokens) It was a really meaty refactor because it affected several files that are like 1,000 to 3,000 lines long which themselves also need refactoring later So I tried 3 approaches to see the difference between: (1) One shot the refactor with Claude 3.7 Max (2) One shot the refactor with Gemini 2.5 Pro (3) Help Agent multi shot the refactor, one sub task at a time and in sequence using Taskmaster Here are my results: 1) First, I used Claude Sonnet 3.7 in Max mode with Thinking enabled. It understood the assignment, but when it attempted to diff in changes, it would also try deleting 500-1,000 lines of code at the same time 🤡 I tried in Ask and Agent mode. It would choke every. single. time. 2) Gemini 2.5 Pro was throttled on Agent mode so I tried in Ask mode This went a little better, except I needed to make changes in core logic, and then apply that change across 20 core function files. It was able to get 3-4 done but by the 5th one it was trying to rename functions and trying to create new core logic that already existed in the app (despite being given those files). Because the context is so much larger, so too were its proposed code changes. Like with Claude, Cursor kept choking when attempting to apply the changes. It was just too much for it to handle at once. 3) Finally, I used my actual current workflow and updated my refactoring task with more details, research and subtasks using Taskmaster The parent task clearly explains how the implementation must go at a low level. What files to touch. What functions to rewrite and import. What rules to follow. And the sub-tasks were the actual individual functions I needed to refactor. Cursor would follow the exact same process for each of the 20 functions using the parent task as the vector and the subtask as the thrust. This… worked flawlessly. 🤯 Instead of trying to get 100% of the task done in each prompt, I requested only 1 of 20 tasks (5%) which it flew through. Since it was on Agent mode and agent has access to Taskmaster, it was able to keep going uninterrupted for about 20 minutes by: - finding the next task and subtask - marking it as in progress - writing a test for the subtask - performing the changes/refactor - validating the new code by running and passing the task’s test - marking the task as done - repeat for the other 19 subtasks I turned on YOLO mode, sent the prompt and went for breakfast with my wife. When I came back, the refactor was 100% done and tested. If you want to build ambitious projects with @cursor_ai, stop twisting its arms and brute forcing your way through context limits There’s a better way now