Can a prompt replace Rezi?

Build a resume linter whose rules are readable and whose score is explainable.

Input is a resume as structured data — reuse the JSON Resume shape rather than inventing one. Output is a score out of 100 and, for every point deducted, the exact line responsible and the rule that fired.

Start with rules that are checkable rather than tasteful, each in its own file with a weight and a one-sentence rationale: every bullet opens with a past-tense action verb; at least sixty percent of bullets contain a number, a percentage or a currency amount; no bullet exceeds two rendered lines; no first-person pronouns; no duplicated opening verb within a section; every role has a date range with no unexplained gap over six months; required sections present. Weights live in one config file, so disagreeing with the scoring means editing a number, not the code.

The report is the product. Print the score, then a list where each row is: the deduction, the rule name, the offending text, and the rationale. Never emit a bare number.

Then add targeting. Given a job description, extract its significant terms — nouns and noun phrases, minus a stop list — and print three columns: terms present in your resume, terms absent, and terms present but only in a section a skim would miss. Absent is a prompt to think, not an instruction to insert; say so in the output, because keyword stuffing is exactly how a resume stops reading like a person wrote it.

With an LLM API key, add a per-bullet rewrite that must preserve every number in the original and is shown as a diff. Assert that in a test: a rewrite that drops or changes a figure fails.

Do not build a template gallery or a PDF renderer here — this tool reads a resume and judges it, and stays small by not also producing one.