# The grading rubric What `scripts/grade_resume.py` measures, where each convention comes from, and how much confidence it deserves. Read this before changing a weight. ## The honest framing Every threshold here is a **convention**, not a measured optimum. Careers services and recruiter guidance converge on these numbers, but almost none of them come from controlled studies with published effect sizes. The most-cited number in the whole field — "recruiters spend 7.4 seconds on a resume" — comes from a [single 2018 study of 30 recruiters](https://www.theladders.com/static/images/basicSite/pdfs/TheLadders-EyeTracking-StudyC2.pdf) and is [openly contested](https://spectacletalentpartners.com/is-the-6-second-resume-scan-a-myth/). So the tool is built to be *useful without being authoritative*: every deduction names the exact line it came from and the convention it applied, so a disagreeing user can see precisely what was judged and overrule it. A score with a visible derivation is a checklist. A score without one is a horoscope. Confidence ratings below: - **High** — near-universal agreement across independent sources, and the check is objective (the text either contains a number or it does not). - **Medium** — broad agreement on direction, arbitrary threshold. - **Low** — plausible, contested, or prone to false positives. Weighted lightly. --- ## 1. Quantified impact — 22 points · confidence High **Convention:** at least half of experience bullets should carry a number, percentage or currency amount. The direction is the least controversial advice in the field: an achievement with a magnitude beats one without. [Yale's careers office](https://ocs.yale.edu/resources/writing-impactful-resume-bullets/) frames the bullet itself as "Accomplished [X] as measured by [Y] by doing [Z]" — measurement is structural, not decorative. The **50% threshold** specifically is a widely repeated rule of thumb ([Resume Worded](https://resumeworded.com/how-to-quantify-resume-key-advice), [Hiration](https://www.hiration.com/blog/resume-critique-rubric-career-centers-higher-ed/)), not a measured optimum — hence Medium confidence on the number even though the principle is High. Deduction scales with the shortfall rather than being all-or-nothing, so a resume at 40% loses a little and one at 0% loses everything. **Detection:** a percentage, a currency symbol followed by a digit, a multiplier (`3x`), a magnitude suffix (`40k`), or any bare number. Deliberately generous — a false negative here is worse than a false positive, because the tool is telling you to go look at a line, not deleting it. ## 2. Strong openings — 18 points · confidence High **Convention:** experience bullets open with a concrete verb, not a duty phrase. "Responsible for the billing system" describes an assignment. "Rebuilt the billing system, cutting settlement time from 4 hours to 12 minutes" describes a contribution. Every source consulted agrees, and the [phrases flagged](https://www.monster.com/career-advice/article/resume-buzzwords-0417) — *responsible for*, *helped with*, *assisted with*, *involved in*, *duties included* — recur across all of them. **Section-scoped, and this matters.** Only experience bullets are judged. A summary line is allowed to be a noun phrase ("Platform engineer with nine years…") and an education entry is a credential, not an achievement. An earlier version of this tool applied the rule everywhere and confidently marked a correct summary and a correct degree line as defects. Scoping is not a nicety; without it the tool is wrong on well-written resumes. Present tense is accepted — it is correct for a role you currently hold. ## 3. Verb variety — 12 points · confidence Medium **Convention:** openers should not repeat heavily; roughly 10–12 distinct verbs across a one-page resume. Widely advised, no hard evidence. The underlying point is real: six bullets all starting "Developed" make six different pieces of work read as one. Weighted below quantification because the threshold is soft and a legitimately repetitive role exists. ## 4. Bullet length — 12 points · confidence Medium **Convention:** roughly 12–32 words. Flagged under 10 or over 40. Guidance clusters at [15–30 words / one to two lines](https://hireflow.net/blog/best-bullet-point-length-for-resume-real-guidance). The mechanism is skimming: a 60-word bullet is a paragraph wearing a bullet's clothes, and the achievement inside it is invisible at a glance. A very short bullet is usually a duty with the result missing. Education entries are exempt — a degree line is *supposed* to be short. ## 5. Clean language — 16 points · confidence High for clichés, Medium for pronouns **Convention:** no self-descriptive clichés, no first person. The cliché list is the intersection of several published lists ([Monster](https://www.monster.com/career-advice/article/resume-buzzwords-0417), [Novoresume](https://novoresume.com/career-blog/resume-buzzwords-to-avoid), [Korn Ferry](https://www.kornferry.com/insights/this-week-in-leadership/5-cliches-to-keep-off-your-resume)): *team player*, *hard worker*, *detail-oriented*, *results-driven*, *self-starter*, *go-getter*, *think outside the box*, and friends. The reasoning is sound and mechanical: these assert a trait rather than evidencing it, and because nearly every applicant writes them they cannot differentiate anyone. Dropping the first person is a formatting convention with less behind it, so it costs fewer points. ## 6. Structure and completeness — 20 points · confidence High Objective presence checks: a reachable email and phone, the four expected sections, a date range on every role, a title under every role header. **The most valuable single check in the tool lives here.** A role header is only recognised when it carries a date range. Without one, the parser cannot distinguish it from a bullet, so the company and job title get swept into the bullet list — and the *formatted document comes out wrong*, with the employer rendered as a bullet point. That is a silent, disqualifying defect, and it is one edit to fix. It is reported with an 8-point deduction and an explicit explanation. --- ## Job match — scored separately, never blended When a job posting is supplied, the tool reports keyword coverage as its **own score**. It is deliberately not folded into the resume score: "well written" and "matches this specific posting" are different questions, and averaging them produces a number that answers neither. **Why keyword coverage at all.** Applicant tracking systems weight the overlap between posting and resume, with terms in the experience section counting for more than elsewhere ([Jobscan](https://www.jobscan.co/blog/top-resume-keywords-boost-resume/)). Published guidance puts a workable target around 75% coverage. **Extraction is prominence-based, not frequency-based.** Ranking a job posting by raw word frequency surfaces *looking*, *ideally*, *comfortable* — a posting is mostly ordinary English. Two cheap signals do far better: - a word capitalised somewhere other than a line or sentence start is almost always a technology or proper noun — Kubernetes, Terraform, Istio, Bazel - a two-word phrase that recurs is usually a real concept — *continuous integration*, *incident response*, *feature flags* Everything else must appear more than once to qualify. On a real platform- engineering posting this cut the term list from 40 mostly-noise words to 21 almost entirely meaningful ones. **Two things it will not do.** It will not tell you that `K8s` and `Kubernetes` are the same thing — matching is literal, and the report says so where the user will read it. And it flags terms used **eight or more times** as possible stuffing, because modern systems penalise unnatural keyword density rather than rewarding it ([Workday's 2026 change](https://www.uppl.ai/ats-resume-keywords) specifically flags it). Advice that only ever says "add more keywords" is advice that eventually gets someone filtered out. The report tells users to add only terms they can honestly back up, in the bullet where they actually did the thing. --- ## Deliberately not measured **Employment gaps.** Technically easy — parse the dates, subtract. Excluded on purpose. Flagging gaps encodes a penalty against career breaks, illness, caring responsibilities and layoffs, and a tool has no business quietly applying that on a user's behalf under the banner of objectivity. This is a values decision, not an oversight, and it is stated in the script's own docstring so it cannot be removed by accident. **Anything inferring demographics** — name origin, graduation year as an age proxy, photographs. Out of scope permanently. **Tense consistency.** Sounds checkable; it is not. Convention disagrees with itself about current roles, and a naive check produces constant false positives on legitimate resumes. **Total page count.** The formatter already reports it, and "too long" depends entirely on seniority and field. See `resume-notes.md` §5. **Whether the writing is *good*.** The tool checks structural properties that correlate with good bullets. It cannot tell whether an achievement is impressive or whether a claim is true. Nothing that reads text can, and a score that implied otherwise would be lying.