Course Content
How to Write a Good Resume
15 sections · 30 lessons
Bias, calibration and reading the rubric backwards
A rubric is necessary but not sufficient. Screeners still bring biases to the page, and two screeners with the same rubric still disagree about what the evidence is worth. This lesson covers the controls for both, and then turns the whole section around to show what it means for the candidate writing the resume.
The common thread is structure. The controls that work do not ask anyone to be a better person; they narrow the room in which judgement can drift. And the resume that does best under that structure is the one this course has been teaching you to write.
Bias in resume screening
Resume screening is one of the most-studied points in hiring, and the findings are consistent and uncomfortable.
Where bias enters
Name. Multiple audit studies across several countries have sent otherwise identical resumes differing only in the name, and found significant differences in callback rates by apparent ethnicity and gender. The effect is well replicated.
School. Institutional prestige acts as a proxy for ability far beyond its predictive value, and it correlates strongly with family income rather than with engineering capability.
Company brand. A candidate from a well-known company is read more generously. Sometimes justified; frequently a shortcut that skips the actual evidence.
Career gaps. Penalised more heavily than performance data supports, and the penalty falls disproportionately on people with caregiving responsibilities.
Writing polish. Non-native speakers of the hiring language get read as less capable on the basis of prose style rather than engineering ability.
Photos, where included — the reason photo-free convention exists in several markets (see Photos, personal data, and regional norms).
Controls that measurably help
| Control | Effect |
|---|---|
| Anonymise name, photo, and address before review | Removes the most-studied bias directly |
| Anonymise institution where feasible | Reduces prestige shortcutting |
| Structured rubric with evidence standards | Reduces the room in which bias operates |
| Score independently before discussion | Prevents anchoring on the first opinion |
| Multiple screeners on borderline candidates | Averages out individual bias |
| Track outcomes by demographic where lawful | You cannot fix what you do not measure |
Anonymisation and structure are the two that do most of the work. Neither requires anyone to be less biased — they narrow where bias can operate, which is why they work.
Calibrating multiple screeners
Bias is one source of noise. Plain disagreement is another. A candidate's outcome should not depend on which screener drew their resume. Without calibration, it substantially does.
The problem
Give the same twenty resumes to three experienced engineers with the same rubric and you will routinely see meaningful disagreement — candidates one screener advances and another declines.
Everyone is reading the same words. They differ on what the words are worth.
Running a calibration session
1. Pick five resumes spanning the range: two clear advances, two clear declines, one genuinely borderline.
2. Everyone scores independently. No discussion, no shared document until all are done.
3. Compare and find the disagreements. Those are the interesting cases.
4. Discuss the evidence, not the conclusion. The productive question is "which line did you score that on?" not "why did you like them?" Disagreements usually resolve into one of two causes:
- Different evidence standards. One person counts "contributed to the migration" as ownership; another does not. Fix the descriptor.
- Different weighting. Both agree on the evidence and disagree on how much it matters. Fix the rubric weights.
5. Update the rubric. The session's output is a changed document, not a shared feeling.
6. Repeat quarterly, and whenever a new screener joins.
Ongoing calibration
- Double-screen a sample. Have two people independently screen 10% of applications and track agreement. A falling rate signals drift.
- Review the declines. Once a quarter, have someone re-screen a random sample of rejected candidates. This is the only way to find false negatives, which are otherwise invisible.
- Close the loop from interviews. If candidates who scored 2 consistently perform well in interviews, your rubric is mis-weighted.
Reading the rubric backwards
Now the reason a candidate should read this section: knowing how a structured screen works tells you exactly what to put on the page.
What survives a structured screen
A structured screen scores against written evidence standards. That means:
| Resume property | Structured screen | Unstructured screen |
|---|---|---|
| Specific outcomes with numbers | Scores directly | Helps |
| Explicit ownership language | Scores directly | Helps |
| Stated scope and scale | Scores directly | Often missed |
| Prestigious employer or school | Neutral to slightly positive | Large effect |
| Attractive design | No effect | Small effect |
| Buzzwords and adjectives | No effect | Small negative |
| Personal rapport signals | Removed by anonymisation | Real effect |
The pattern is clear: structured screening rewards evidence and neutralises everything else.
What this means for you as a candidate
1. Evidence is the only reliable strategy. It scores in both kinds of screen. Charm, design, and pedigree work in one and not the other.
2. Make the evidence quotable. A screener records "score 3 — 'owned routing service, 12M requests/day, on-call rotation of 6'". Write lines that can be quoted into a scorecard as they stand. This is the underlying reason context lines and numbers work so well.
3. Map yourself to the likely criteria. The job description's top three to five requirements are approximately the rubric. That is exactly what the job-description diff does — you were reverse-engineering the rubric without it being named.
4. Do not rely on being remembered. In a structured process, nothing survives that was not written into a scorecard. If your best quality is not on the page in a quotable form, it does not exist.