Scoring 500 resumes at 92% accuracy — what I learned about prompts

The resume analyzer hit 92% accuracy on a hand-labelled set of 500+ resumes. That number came after a long stretch of lower numbers, and the gap between the two states was almost entirely about how the prompt handled bad input.

A resume is a hostile document. The PDF has columns, headers, footers, multi-page tables, scanned images with OCR errors, and a creative freedom that no structured schema can predict. Most LLM evaluation runs on these inputs are dominated by a small number of edge cases that the prompt fails on. The fix is rarely “a better model.” It is almost always “stop asking the model to do things that the parser should have done.”

Three lessons from the production run. First, validate the parser output before you send it to the model; do not let the model see raw OCR text. Second, score with a rubric, not a single number; the rubric gives you something to debug when accuracy drops. Third, run the prompt against the labelled set on every change; the 92% number is the result of dozens of small wins, not one prompt rewrite.