Hands-on test: AI skips step-by-step orders unless they leave a trace in NOTES.md

The author started from a common belief that important instructions get followed if they are short and placed prominently, pointing to Claude Code's official documentation about bloated instruction files. To test that belief properly, they built a task with known answers, using 12 log-like files where the right value had to be pulled from lines starting with RESULT: and a correct answer sheet generated by script and kept outside the directory given to the subagent.

Across the first two experiments, the amount of material (12 vs. 60 files) and the number of lines in a prompt did not affect follow-through, and the subagent's tool-call counts for the one-at-a-time instruction stayed low. This pointed to a different variable: whether following the instruction left any trace in the graded output. That led to the third experiment, where two instructions with the same position, scope and required effort were placed adjacent to each other, differing only in whether compliance left a record, and the trace-leaving instruction was met 4 out of 4.

The author is candid about the limits of the evidence, noting that the all-perfect answer scores mean the task may have been too easy to draw strong conclusions, and that the reason for the improvement when A was added is not explained by this design. The practical takeaway offered is not to rewrite the instruction but to add a nearby line that asks for an output showing the process, or to treat an untraceable instruction as one that may not be followed.