Claude Code does not follow user instructions consistently

Issue

Claude Code repeatedly fails to follow explicit user instructions during extended sessions. Specific problems observed:

1. Does not read data before acting — User repeatedly has to say "read first, then act." Claude rushes to write code or make comparisons without understanding the source data.

1. Makes assumptions instead of investigating — Assumes column names, field meanings, function signatures instead of reading them. Gets corrected, acknowledges, then does it again minutes later.

1. Does not listen to what the user actually says — Pattern-matches to its own context instead of processing the user's actual words. When asked "where is the app path" responds about a test file it just wrote.

1. Repeats the same mistakes after correction — User says "don't rush" multiple times. Claude acknowledges every time. Continues rushing. The acknowledgment is performative, not functional.

1. Does its own calculations when told not to — In a payroll verification context, was explicitly told "the engine calculates, you compare." Proceeded to do manual arithmetic and got it wrong.

1. Fails to follow explicit workflow instructions — Told to verify company-by-company, investigate all tables first, present findings before acting. Repeatedly skips steps and jumps ahead.

These are not edge cases. This happened continuously across a multi-hour session despite repeated, explicit corrections.

Submitted by user via Claude Code session.