The Preferred Candidate
A practicum in career readiness, applications and selection: take one real vacancy you could realistically pursue, establish whether you are ready for it, and build an application and a performance that a panel could check.
For job-seekers with one real vacancy, at one named organisation, that they could realistically pursue.
- How long
- About 14 weeks at 3 to 4 hours a week
- Cost today
- Level 0 free
- The rest
- Early access, by invitation or request
By the end
- Answer one real vacancy with an application backed by evidence someone other than you will vouch for.
- Map what you bring against a named vacancy, checked by a second person.
- Turn a posting into a ranked list of what is actually being assessed.
- Build a master CV and tailor it without changing a fact.
- Defend what you wrote under structured questioning.
Most job-seeking advice is about appearing qualified. It teaches formatting, phrasing, confidence and the vocabulary of achievement. It works on the document and leaves the candidate untouched, which is why people who follow it well still lose to people who followed it badly and had better evidence.
This course starts one step earlier, with a question that is harder and more useful: what would make an employer see you as the preferred candidate for this role, and is any of that currently true of you?
Everything here runs against one real vacancy at one named organisation. Not an example vacancy. Not the kind of role you are interested in. One posting, with a closing date, that you could realistically pursue. You will assess your readiness against it, translate what it is actually asking for, attach evidence to each requirement, obtain at least one piece of evidence that somebody other than you will vouch for, build a master CV and a tailored version, write an application that could not have been sent anywhere else, and defend the whole thing under structured questioning.
What the evidence says about courses like this one
The honest position is that structured job-search programmes work, that the effect is real and moderate rather than transformative, and that the design details decide whether you get any of it.
The reference figure comes from a meta-analysis of 47 experimentally or quasi-experimentally evaluated job-search interventions. Treated participants had odds of obtaining employment 2.67 times those of controls (Liu, Huang & Wang, 2014). That is an odds ratio and not a probability ratio: against a baseline employment rate of 30 per cent it moves the rate to roughly 53 per cent, which is substantial and is not the same thing as making you 2.67 times more likely to be hired.
The same paper carries the condition that makes it usable. Interventions promoted employment only when they combined skill development with motivation enhancement. Programmes that taught technique alone did not move employment. Neither did programmes that raised confidence alone.
Effects in Liu et al. were weaker for the long-term unemployed than for the recently unemployed, and weaker for middle-aged participants than for younger and older ones.
If you have been out of work for a long time, this course is still worth running and the evidence says it will do less for you than it does for someone three weeks into a search. Level 2 is built for exactly that case, because the interventions that work best for the least advantaged are the ones that produce verifiable third-party evidence.
The result the whole course is organised around
Two field experiments, in two African labour markets, isolate the thing that actually moves hiring outcomes.
In Johannesburg, 6,891 job-seekers took a battery of skill assessments. One arm received their results as a branded certificate they could credibly share with an employer. Another received the same results privately, in a form no employer could attribute to a third party. The public certificate raised employment by 5.2 percentage points against a control mean of 30.9 per cent, and weekly earnings by 34 per cent. The private certificate produced close to zero employment effect (Carranza, Garlick, Orkin & Rankin, 2022).
Same information. Same person. Same skills. The difference was whether anybody other than the candidate would vouch for it.
In Addis Ababa, 3,052 job-seekers were randomised into a transport subsidy costing 19.80 US dollars per person, a two-day job application workshop costing 18.20 dollars per person, or a control group. At eight months the workshop had raised permanent employment by 6.9 percentage points against a control mean of 12.0 per cent, against 3.3 points for the subsidy. At four years the workshop was still raising earnings by about 25 per cent and the transport subsidy's effect had disappeared entirely (Abebe, Caria, Fafchamps, Falco, Franklin & Quinn, 2021).
The workshop's Day 1 was aptitude and work-sample tests producing a certificate. Day 2 was CV writing, cover letters and interview technique. That is the syllabus of this course, and the number worth carrying is not the employment rate. It is that among low-educated participants, applications required per permanent job offer fell from 10.5 to 4.6.
The workshop did not work by making people apply more. It worked by making each application more than twice as likely to convert.
Every hour you are about to spend goes into the conversion rate of a small number of serious applications. If your current strategy is volume, this course will feel slow for six weeks and then stop feeling slow.
The six plateaus
You move through six capabilities, each gated by behaviour someone else could watch you fail.
First you fix one real vacancy and establish, against an external check rather than your own impression, what you currently bring to it. Then you take the posting apart and work out what is actually being assessed, which is rarely what is listed first. Then you attach evidence to each requirement and obtain at least one piece of evidence a third party will confirm. Then you build a master CV and learn to tailor it without changing a fact. Then you write the application, and learn which parts of that advice are evidenced and which are convention. Finally you enter structured selection and defend what you wrote, in front of people whose job is to find the seam.
Cadence
Plan for three to four hours a week: about an hour on lessons and checks, two hours on the exercises against your real vacancy, and thirty minutes on the record.
Fourteen weeks. Levels 0, 1, 2 and 4 take two weeks each. Level 3 takes three, because a master CV built properly is a week of recall work before it is a document. Level 5 takes three, because it contains conversations with real people that you will want to postpone.
The source programme this course is built from runs eight facilitated weeks. This runs fourteen because a session can be attended and a request for a reference letter cannot.
Where the evidence comes from, and where it does not
Almost every psychological finding in this course was established on a Western sample. In an audit of the field, 96 per cent of research subjects came from Western industrialised countries holding 12 per cent of the world's population, and 68 per cent came from the United States alone (Henrich, Heine & Norenzayan, 2010).
For this subject the gap is sharper than usual and it runs in a specific direction. A quasi-exhaustive register of correspondence experiments worldwide between 2005 and 2020, covering more than 900,000 fictitious job applications, contains no sub-Saharan African country (Lippens, Vermeiren & Baert, 2023). There is no correspondence study, audit study or CV-based field experiment on hiring discrimination in the Nigerian labour market. There is no randomised labour-market field experiment in Nigeria at all.
The strongest evidence in this course about what works comes from South Africa, Ethiopia and Uganda, and it is genuinely African evidence from randomised trials in economies with similar informality and similar screening problems. Use it.
The evidence about CV formatting, callback rates and screening behaviour comes from the United States and Western Europe. Some of it will transfer and some of it will not, and nobody has measured which.
Where your own record and this course disagree, the record is the evidence and this course is the hypothesis.
What this course does not claim
It does not claim that a better application defeats a labour market. Nigeria's headline unemployment rate is around 4 to 5 per cent, which is correct under the international standard and close to useless as a measure of job scarcity, because one hour of paid work in the reference week counts as employment. The figures that describe the actual market are that 93.0 per cent of employment is informal, 14.4 per cent is wage employment, and the broad measure of labour underutilisation is 18.5 per cent (National Bureau of Statistics, Q2 2024; ILO, November 2024). No application improves those.
It does not claim that improving your CV helps everybody equally. In the foundational correspondence experiment, upgrading a résumé raised callbacks by about 27 per cent for applicants with White-sounding names and produced a smaller, statistically undetectable gain for applicants with African-American-sounding names (Bertrand & Mullainathan, 2004). Anyone who teaches "improve your CV" as a universal lever is describing a result that is not universal.
And it does not claim that any of this is a substitute for being able to do the job. The course's own facilitator principle is the correction: if the evidence is weak, improve the capability. If the evidence exists, communicate it better.
A level is passed when you have done something with a real vacancy, a real document or a real person, and can show what happened.
Agreeing with a lesson costs nothing. This page has changed nothing about your candidacy, and it is not supposed to.
The levels
- 0Establish what you actually bringReplace your impression of your own readiness with a map against one named vacancy that a second person has checked.Free
- 1Read the roleTurn one posting into a ranked list of what is actually being assessed, in language the person who wrote it would recognise.Locked
- 2Prove the matchAttach a dated instance to every ranked requirement, and obtain at least one piece of evidence somebody other than you will confirm.Locked
- 3Build the documentOne master CV holding everything, two tailored versions built from it, and a professional story a stranger can follow in ninety seconds.Locked
- 4Write the application that gets readOne application in which every paragraph answers a named criterion, and nothing in it could have been sent to another employer.
Your target role
One real vacancy, one named organisation, one person who will read what you write and tell you when it is thin.
Before you finish Level 0 you fix four things. They stay fixed for fourteen weeks.
One: the vacancy
A real, currently advertised or recently advertised role at a named organisation, with a written description you can open. Not a job title. Not a sector. One posting.
It has to satisfy three tests. You could realistically pursue it, meaning it is not two grades above anything you have held and not in a country where you have no right to work. You want it enough to spend fourteen weeks on it. And its description runs to more than four lines, because a four-line posting cannot be taken apart.
If the exact posting closes during the course, keep it. You are not required to submit; you are required to build an application that would survive submission. Most learners find a second live posting for the same role type by Level 4 and apply to that one.
If your market advertises poorly, which is the common case for senior and for informal-sector roles, use the closest full posting you can find for the same role at a comparable organisation, and note the substitution in writing.
Two: the second vacancy
One further real posting for a different role that you could also realistically pursue. You will use it once, in Level 3, to produce two tailored CVs from one master document and see what changes.
Choose it now, while you are not yet invested in the first one.
Three: the reader
One person who will read what you produce and tell you when it is evasive. Not a mentor and not a cheerleader. Somebody whose reaction to a weak sentence is to say so.
They do not need to work in your field. Three of the six exercises specifically require somebody who does not, because a reader inside your profession fills in your gaps automatically and a reader outside it cannot.
Name them before Level 0 closes, in writing, with the date you asked them.
Four: the third party
One person who could credibly vouch for something you claim: a former manager, a client, a lecturer who supervised your project, the chair of the association you served on. Somebody with a name, a title and a working email address.
You will approach them in Level 2. Identify them now, because the single largest finding in this course is about the difference between evidence you assert and evidence somebody else confirms.
"Programme Officer, Monitoring and Evaluation, at a health NGO in Abuja. Posting downloaded 4 March, closes 28 March, four pages, eleven essential requirements. Second vacancy: Data Analyst at a commercial bank, Lagos. My reader is Ada, who runs a logistics business and knows nothing about M&E. My third party is Mr Okonkwo, who was my supervisor at the state ministry from 2022 to 2024, and I have his current email."
"I am looking for a good role in development or possibly in tech, ideally something remote, where I can use my skills and grow." No posting, no organisation, no requirements, no dates, no people, and nothing that could be assessed as met or unmet.
If two postings come to mind and one of them would require you to write down a requirement you do not meet, that is the one with the information in it.
The other one is the role you already have, described differently.
Rules of practice
Seven rules. Four are supported by evidence, three are design choices, and the difference is marked.
1. Everything runs against one named vacancy
Design choice. Not a sector, not a role type, not "opportunities in consulting". The reason is in the evidence on what actually moved outcomes: the interventions that worked produced something attributable to a specific candidate and legible to a specific employer (Carranza et al., 2022; Abel, Burger & Piraino, 2020). A general improvement is not attributable to anything.
2. Every claim carries a locator
Design choice, with support by analogy. A locator is a date, a name, a number, a document somebody could open. The support comes from goal monitoring rather than from hiring: prompting people to monitor progress raised attainment by d = 0.40 across 138 trials and 19,951 people, and physically recording progress outperformed tracking it without a record, 0.43 against 0.29 (Harkin et al., 2016). That literature is about personal goals. What transfers is narrow and sufficient: a claim without a locator is an account written by the person with the strongest interest in the outcome.
3. Third-party evidence outranks self-authored evidence
Evidence: strong. A skills certificate that could be credibly shared with employers raised employment by 5.2 percentage points and earnings by 34 per cent; the identical assessment delivered privately produced close to zero employment effect (Carranza, Garlick, Orkin & Rankin, 2022). A reference letter raised employer callbacks from 4.2 to 6.7 per cent, with the entire gain accruing to women (Abel, Burger & Piraino, 2020). Nothing else in this literature has that shape.
4. Fix the mechanics before you fix the message
Evidence: strong. Five spelling errors on a résumé cut interview probability by 18.5 percentage points, and two errors by 7.3 points, across 445 recruiters and 1,335 vignettes (Sterkens et al., 2023). In a separate study of 512 French recruiters, spelling errors raised the odds of rejection by 3.65 and did as much damage as lacking professional experience (Martin-Lacroux & Lacroux, 2017). In a randomised trial with 480,948 job-seekers, purely mechanical writing suggestions with no content generation raised hires by 8 per cent (Wiles, Munyikwa & Horton, 2023). This is the best-evidenced piece of application advice in existence and it is the least interesting one.
5. Where a rubric is published, answer it line by line
Design choice, with a documented mechanism. The clearest public example is the UK Civil Service Success Profiles system, where each named behaviour is scored 1 to 7 against the evidence offered and evidence not offered against a named criterion cannot be scored at all. No published reliability or validity study of Success Profiles exists, which is worth knowing. The argument for answering the rubric is not that the rubric is validated. It is that unclaimed evidence is unscoreable.
6. You do not fake, because faking does not do what people think
Evidence: moderate, and more interesting than the rule sounds. Over 90 per cent of candidates use some impression-management tactic and about 75 per cent invent an example in a real interview (Levashina & Campion, 2007). Interviewers detect faking at 52.4 per cent against a chance rate of 50 (Reinhard et al., 2013). So faking is common and undetected. The reason not to do it is what it costs: in a within-subjects experiment, faking raised interview scores by d = 1.49 and simultaneously destroyed the interview's ability to predict citizenship behaviour, dropping that correlation from .51 to .13 (Buehl, Melchers, Macan & Kühnel, 2019). Across 27 studies, deceptive impression management has close to zero average association with interview evaluations, while honest impression management is associated with better ones. You are buying a score you cannot sustain in the job.
7. You claim your own gates
Design choice. Nobody marks these for you. Marking one early costs only you.
One lesson block read, at least one knowledge check answered wrong, the exercise run against the real posting, three dated lines in the record, and one thing you avoided named honestly with the reason you avoided it.
All lessons read, all checks correct first time, the exercise done in your head about the kind of role you want, the record blank, and a private sense that you already do most of this.
Assessment
What counts, what does not, and what you should hold at the end.
What is assessed
| Component | Weight | The evidence |
|---|---|---|
| Readiness map and its external check | 15% | The map against one named vacancy, and what a second person changed in it |
| Role analysis and evidence map | 15% | Every essential requirement ranked, with a dated instance or a marked gap against each |
| Third-party evidence obtained | 10% | One reference letter, certificate, assessment result or written confirmation, with the date you asked |
| Master CV and two tailored versions | 20% | The master document, two tailored CVs, and the diff between them |
| The application | 15% | A supporting statement or cover letter in which every paragraph names a published criterion |
| The Selection Room | 25% | A full structured interview and panel, with what you were asked, what you answered and what you could not defend |
What is not assessed
Confidence is not assessed. Fluency in describing your career is not assessed. Whether you got the job is not assessed, because that outcome is determined mostly by things you do not control, and a course that graded on it would be grading on the labour market.
A measurement warning that applies to the first thing you will do
The course opens by asking you to rate your own readiness, and self-ratings are the least reliable instrument in it.
Two independent syntheses thirty-two years apart converge on the same number. Across 55 studies, self-evaluation of ability correlated r = .29 with performance measures (Mabe & West, 1982). Across 22 meta-analyses spanning academic, intelligence, language, medical, sports and vocational domains, the mean was r = .29 again, with individual effects ranging from .09 to .63 (Zell & Krizan, 2014). Self-assessment shares roughly 8 per cent of its variance with actual ability.
The familiar story about this, that incompetent people are wildly overconfident, is largely a statistical artefact. Plotting the difference between self-rating and score against the score itself generates the Dunning-Kruger pattern from pure random numbers (Nuhfer et al., 2016, 2017). Applied to 929 community participants with an objective ability measure, the two statistically valid tests of the hypothesis both failed to find it (Gignac & Zajenkowski, 2020). A double-censored model reproduced the whole pattern from scale boundary effects alone in 665 students (Magnus & Peresetsky, 2022).
So the correct statement is not that you are probably deluded. It is that your self-rating tracks reality weakly, and that Mabe & West identified the conditions under which it tracks better: when the rater expects the rating to be checked against an objective criterion, when the rater has practised self-evaluation before, and when the instruction asks for comparison against peers rather than against an absolute standard. The Level 0 tool is built to all three conditions, which is why it asks who will check each line.
The check
Ten statements, each scored one to five, run before Level 0 and again after Level 5. Run it now and record the date. The total carries almost no information. The change on individual lines carries it, particularly the lines on interpreting a job description, connecting requirements to evidence, and defending a claim under questioning.
What you should hold at the end
A readiness map against one named vacancy, with an external check recorded. A ranked translation of every requirement in that posting. A Role Evidence Map with a dated instance or a marked gap against each. At least one piece of third-party evidence you did not write. A master CV. Two tailored CVs from it, with what changed and what did not. One application in which every paragraph answers a published criterion. An Interview Evidence Bank of at least twelve situations. A record of one full structured interview and one panel, including the question you could not answer.
Sources
Every figure in this course, with the study behind it and what it can and cannot support.
Full citations with links are in SOURCES.md in the course directory. This page is the map: what each level rests on, and how strong it is.
How to read the grades
Strong means large samples, randomised or field-experimental designs, and either replication or a bias-corrected estimate. Moderate means well conducted but single-source, or meta-analytic with known limits. Weak means small samples, one laboratory, a commercial survey, or a claim that circulates without a controlled test behind it. Vendor means the figure was produced by a party selling the remedy.
Where a figure is weak, the lesson says so rather than this page.
A structural warning about this subject's evidence base
Almost the entire public stock of job-search statistics is circular. Vendor A cites a news article that cited vendor B that cited nothing. The most-cited academic report on automated screening sources its own headline prevalence figure to a résumé-optimisation company. Several of the numbers most confidently repeated in careers training have no study behind them at all, and this course names each one rather than quietly omitting it.
Level 0, on readiness and self-assessment
| Finding | Figure | Grade |
|---|---|---|
| Self-evaluation of ability correlates weakly with measured ability | r = .29, SD = .25, 55 studies (Mabe & West, 1982); r = .29, SD = .11, 22 meta-analyses (Zell & Krizan, 2014) | Strong, two independent syntheses |
| The Dunning-Kruger curve is largely a plotting artefact | Pattern reproduced from random numbers, N = 1,154 (Nuhfer et al., 2016, 2017); both valid tests null, N = 929 (Gignac & Zajenkowski, 2020); reproduced from scale boundaries, N = 665 (Magnus & Peresetsky, 2022) | Strong |
| Self-assessment improves under specific conditions | Nine measurement conditions accounted for R = .64 of variance in validity coefficients (Mabe & West, 1982) | Strong |
| Years of experience barely predicts job performance | ρ = .07, lower 80% credibility value −.07 (Sackett, Zhang, Berry & Lievens, 2022) | Strong |
| People apply at roughly half the stated qualifications, and the gender gap runs the other way | Men 52.1%, women 55.7%; N = 10,468 online randomised trial (Nicks et al., Behavioural Insights Team, 2022) | Strong |
| "Men apply at 60% of qualifications, women at 100%" | Traced to an unnamed, unpublished Hewlett-Packard internal report cited second-hand in three books; no primary document exists | Debunked |
| Nigeria's headline unemployment rate changed for definitional reasons | Same 2011 data: 2.2% under the one-hour ILO definition, 23.9% under the 40-hour NBS definition (NBS Unemployment Committee Review Report, 2014) | Strong |
| The figures that describe the Nigerian market | Informal employment 93.0%, wage employment 14.4%, Q2 2024 (NBS); labour underutilisation LU4 18.5%, Q3 2023 (ILO) | Strong |
Level 1, on reading the role
| Finding | Figure | Grade |
|---|---|---|
| Employers rarely write the posting they screen against | 72% reused or slightly modified an existing posting; 8% wrote a new description; N = 2,275 executives (Fuller et al., 2021) | Strong for the survey, self-report |
| Stated requirements are not binding on the employer's own hires | 47% of leaders said only half or fewer of their middle-skills hires met all listed requirements; 21% said all high-skills hires did | Strong for the survey |
| Employers believe their own filters exclude qualified people | 88% for high-skills roles, 94% for middle-skills (Fuller et al., 2021) | Moderate, employer opinion about their own process |
| Degree requirements are systematically inflated | Production supervisors: 67% of postings required a degree against 16% of incumbents holding one; 26 million postings analysed (Fuller & Raman, 2017) | Moderate, corporate white paper |
| The announced move to skills-based hiring largely did not happen | +3.5 percentage points non-degree hires in changed roles; 37% of firms followed through, 45% "in name only", 18% backsliding; ~97,000 workers a year against 77 million hires; 11,332 roles, 2014 to 2023 (Sigelman, Fuller & Martin, 2024) | Moderate to strong |
| Structured criterion-anchored scoring is a documented public system | UK Civil Service Success Profiles; COPFS publishes the 1 to 7 scale used at sift and interview | Documented system, not a research finding |
| Success Profiles has been evaluated | No published reliability, validity or evaluation study located, from the Cabinet Office, Civil Service Commission, NAO or academic literature | Evidence gap |
Level 2, on evidence
| Finding | Figure | Grade |
|---|---|---|
| A credibly shareable skills certificate raises employment; the same result privately does not | Public certificate +5.2 pp employment on a 30.9% control mean, +34% weekly earnings, +20% hourly wage; private certificate near zero employment effect; 6,891 assessed workseekers, South Africa (Carranza, Garlick, Orkin & Rankin, 2022) | Strong |
| A reference letter raises callbacks, and the gain goes to women | 4.2% to 6.7% employer response, a 60% increase; women's interview invitations +13 pp on a 53% base, employment +5.9 pp on an 11.7% base; men not significant; three trials, South Africa (Abel, Burger & Piraino, 2020) | Strong |
| A two-day application workshop beat a transport subsidy of the same cost, and kept working | Permanent employment +6.9 pp against a 12.0% control mean, against +3.3 pp for the subsidy; workshop still raising earnings ~25% at four years, subsidy dissipated; N = 3,052, Addis Ababa (Abebe et al., 2021) | Strong |
| The mechanism was conversion, not volume | Applications per permanent job offer among low-educated participants fell from 10.5 to 4.6 (Abebe et al., 2021) | Strong |
| Access to employers is not the binding constraint | Two job fairs, 606 job-seekers meeting 170 firms, produced 76 offers to 45 people and 14 acceptances, with no significant employment effect; N = 1,829 job-seekers and 498 firms (Abebe et al., World Bank WP 8092, 2017) | Strong |
| Soft-skill certification changes matching and earnings more clearly than it changes employment | High-ability managers revised skill assessments upward, low-ability managers did not; overall employment effects unclear; Uganda (Bassi & Nansamba, 2022) | Strong paper, direction only reported here |
| References as employers currently use them carry little information | 60.27% of applicants rated "excellent" and 0.23% below average across 6,854 assessments; validity ρ ≈ .29 (Aamodt & Williams, 2005) | Moderate |
Level 3, on the story and the CV
| Finding | Figure | Grade |
|---|---|---|
| Screening takes about 14 to 16 seconds on a first pass | 16.24 s, SD 10.17; 149 participants, 5,960 résumés, calibrated eye-tracking (Lahey & Oxley, 2021); ~14.3 s, 40 real recruiters (Osanami Törngren et al., 2024) | Strong |
| Looking at a CV for longer does not mean rating it higher | Dwell time did not predict rating; hypothesis rejected outright (Osanami Törngren et al., 2024) | Strong, N = 40 |
| "Recruiters spend 6 or 7.4 seconds on your CV" | Two unreplicated vendor eye-tracking studies, one with 30 recruiters, one with sample size undisclosed, neither peer reviewed, both by a company selling résumé services | Weak, and routinely misreported |
| Showing years worked instead of employment dates raises callbacks | +15% for CVs with gaps, +8% without; preregistered field audit, n = 9,022 real applications, plus three lab studies, n = 2,650 (Kristal, Nicks, Gloor & Hauser, 2023) | Strong |
| Past unemployment is not penalised; long current unemployment is | No significant penalty for past unemployment; nine months of current unemployment cost 3.9 pp in medium and low-skill roles and was not significant in high-skill roles; N = 8,466, Sweden (Eriksson & Rooth, 2014) | Strong |
| The penalty starts late and then plateaus | No callback difference at 0, 4, 12 or 24 weeks; 11.0% at 0 weeks against 9.1% at 52 weeks; N = 8,488, US white-collar (Farber, Herbst, Silverman & von Wachter, 2018) | Strong |
| "Take any job to fill the gap" is contradicted | Unemployed applicants 12.6% callbacks against 10.8% with a high-quality interim job and 10.2% with a low-quality one; for high-skill roles 13.4% against ~10%, significant (Farber et al., 2018) | Strong |
| The job-hopping penalty is occupation-specific | Pooled 1.6 pp and not significant; 3.5 pp on a 9.1% base in non-technical roles; reversed for software testers; N = 3,878 (Hollister, Denier & St-Denis, 2025) | Strong |
| A listed technical skills section had no average effect | +1.0 GPA point worth +2.1 to +2.2 Likert points; top internship +0.9 to +1.0; listed technical skills no effect; 72 employers, 2,880 ratings (Kessler, Low & Sullivan, 2019) | Strong, graduate hiring only |
| A photograph is a large measurable penalty in markets where photos are customary | German name bareheaded 18.8%, Turkish name bareheaded 13.5%, Turkish name with headscarf 4.2%; N = 1,474 (Weichselbaumer, 2020). Phenotype callback gap up to 14.5 pp; N = 12,783 (Polavieja et al., 2023) | Strong |
| "Your CV must be one page" | No research origin located. The only study on length, run by a vendor selling CV writing, found the opposite | No evidence either way |
| Improving a CV does not help everyone equally | Résumé quality upgrade raised callbacks 27% for White-sounding names, 8% and not significant for African-American-sounding names; ~5,000 applications (Bertrand & Mullainathan, 2004) | Strong |
| Racial callback gaps did not decline over 25 years | Whites receive 36% more callbacks than African Americans and 24% more than Latinos; 28 studies, 55,842 applications (Quillian, Pager, Hexel & Midtbøen, 2017) | Strong |
| Discrimination is concentrated, not uniform | Average penalty 2.1 pp; the top quintile of firms accounts for about half of all lost contacts; 23 of 108 firms individually identified; ~84,000 applications (Kline, Rose & Walters, 2022) | Strong |
Level 4, on the application
| Finding | Figure | Grade |
|---|---|---|
| Spelling errors carry a large causal penalty | 5 errors −18.5 pp interview probability, 2 errors −7.3 pp; 445 recruiters, 1,335 vignettes (Sterkens et al., 2023) | Strong |
| Spelling errors do as much damage as lacking experience | Odds ratio 3.65 for errors against 2.71 for inexperience; rejection 18.0% for an error-free experienced CV against 36.1% with five errors; N = 512 French recruiters (Martin-Lacroux & Lacroux, 2017) | Strong |
| Mechanical writing improvement raises hires | +8% hires, 95% CI 3% to 13%; wages $19.01 against $17.25 per hour; no effect on applications sent or interviews; 480,948 job-seekers randomised (Wiles, Munyikwa & Horton, 2023) | Strong |
| There is no credible causal estimate of whether a cover letter helps | No published peer-reviewed randomised field experiment isolates presence against absence. The single field experiment is vendor-run, publishes no per-arm rates, group sizes, significance tests or confidence intervals, and carries a publication date preceding the end of its own data collection | Evidence gap |
| Tailoring has one causal estimate, and its value is depreciating | AI-assisted tailoring raised callback probability by 3.56 pp for users, while the correlation between cover-letter quality and hiring fell 51 per cent as tailoring became cheap; 5.5 million bids, ~106,700 jobs (Cui, Dias & Ye, 2025) | Moderate, working paper |
| "75% of résumés are rejected by software before a human sees them" | Traced to Preptel, a defunct résumé-optimisation vendor, via a 2012 trade article. No methodology, sample or dataset was ever published. The original wording was about reduced chances of interview, not automated rejection | Debunked |
| What automated rejection actually keys on | Greenhouse auto-reject is configurable only on Yes/No, single-select and multi-select application form questions, never on résumé text; Jobscan states an ATS "rarely rejects a candidate automatically" | Primary vendor documentation |
| The filtering that does exclude people is human-configured | 94% of employers use their system to filter or rank middle-skills candidates and 92% high-skills; 48% filter on employment gaps over six months (Fuller et al., 2021) | Strong for the survey |
| Free-form personal statements predict little | No incremental validity once test scores and prior grades are controlled (Murphy, Klieger, Borneman & Kuncel, 2009); among the least effective and least fair methods in medical education, 194 articles reviewed (Patterson et al., 2016) | Moderate to strong |
| Employers are redesigning around candidate use of generative AI | 61% of employers reported candidates using AI during interviews without disclosure; 33% redesigned selection processes, up from 23%; 45% gave applicants no guidance; 155 employers, 1.8 million applications (Institute of Student Employers, 2025) | Moderate |
Level 5, on the room
| Finding | Figure | Grade |
|---|---|---|
| The structured interview is now the highest-validity selection method known | ρ = .42, SD .19, 80% credibility interval .18 to .66; unstructured .19 with a lower bound of −.01; GMA .31; work samples .33; years of experience .07 (Sackett, Zhang, Berry & Lievens, 2022) | Strong |
| The older figures everybody quotes were inflated by a correction error | Correction factors derived from predictive studies were applied to a database roughly 80% concurrent; structured interviews fell .51 to .42, unstructured .38 to .19, work samples .54 to .33 | Strong |
| Validity rises with structure and then plateaus | .20, .35, .56, .57 across four structure levels; 15 to 33 studies per level (Huffcutt & Arthur, 1994) | Moderate, entry-level jobs, pre-2022 corrections |
| Past-behaviour questions out-predict situational ones, and the rubric matters more than the question | Past behaviour ρ = .56 against situational .45; past behaviour with descriptively anchored rating scales .63 against .47 without; k = 49 combined (Taylor & Small, 2002) | Moderate, relative comparison only |
| At high job complexity the situational question collapses and the past-behaviour question holds | Situational .30 against behaviour description .51 at high complexity; k = 6 and k = 8 in the pivotal cells (Huffcutt, Conway, Roth & Klehe, 2004) | Moderate, thin cells, contradicted on moderation by Taylor & Small |
| What the interview partly measures is whether you worked out what it was measuring | Ability to identify criteria correlated .23 with interview performance and .29 with job performance; controlling it dropped the interview-performance link from .24 to .19 and non-significant; N = 97 (Ingold et al., 2015) | Moderate |
| "The decision is made in the first seven seconds" | No primary source exists. In a field study of 166 interviewers and 691 applicants, 4.9% reported deciding within one minute, 69.9% after five minutes or later, and 22.5% only after the interview ended (Frieder, Van Iddekinge & Raymark, 2016) | Debunked |
| Rapport before the questions does carry into the score | Initial competence rating correlated .42 with the structured interview score from the same interviewer and .25 and .24 from different interviewers; initial impression correlated .22 with a real internship offer; N = 189 (Barrick, Swider & Stewart, 2010) | Strong |
| "93% of communication is nonverbal" | Assembled from two 1967 studies using single spoken words against still photographs, female participants only, like and dislike judgements only. In the thin-slice meta-analysis, transcripts alone reached r = .29 against tone of voice at .26 | Debunked |
| Anxious candidates think they did far worse than the interviewer thought | Anxiety correlated −.15 to −.49 with self-rated performance against −.07 to −.28 with interviewer-rated performance (McCarthy & Goffin, 2004); γ = −.03 in 8,343 real applicants across 373 companies (McCarthy et al., 2021) | Strong |
| Interview training moves interview scores | d = 0.60 on performance and 0.55 on hireability from one virtual training session against an active control; randomised, N = 70 (Langer, König, Gebhard & André, 2016) | Moderate, student sample |
| Simulation-based interview training moves real employment outcomes | Job offers at six months 47.8% against 14.3%, OR 8.73, 95% CI 1.17 to 65.00 (Smith et al., 2015); competitive employment at nine months 43% against 28%, adjusted OR 3.03, N = 90 (Smith et al., 2022) | Moderate, clinical populations, wide intervals |
| Faking raises your score and costs you the thing the interview was measuring | Faking raised scores d = 1.49; validity for citizenship behaviour fell from .51 to .13; N = 111 within-subjects (Buehl et al., 2019) | Moderate |
On negotiation and on how jobs are found
| Finding | Figure | Grade |
|---|---|---|
| Ambiguity about whether negotiation is permitted creates most of the gender gap in initiating it | Where negotiability was ambiguous, men 10.6% and women 8.2%; where the ad said the wage was negotiable, men 22.0% and women 23.9%, reversing the gap; overall initiation 9.4% to 23.1%; N = 2,422 (Leibbrandt & List, 2015) | Strong |
| The gender gap in negotiation outcomes is small and shrinks with three specific inputs | Hedges' g = 0.20 favouring men; moderators: prior negotiation experience −0.26, information about the bargaining range −0.25, negotiating on behalf of another −0.31; 51 studies, N = 10,888 (Mazei et al., 2015) | Strong |
| "57% of men and 7% of women negotiate their starting salary" | Small single-cohort convenience sample of one university's master's graduates, not designed as a population estimate, never replicated at that magnitude | Weak |
| Weak ties cause more job mobility than strong ties, but the relationship is an inverted U | Ties with more than ten mutual connections reduced job transmission probability; >20 million subjects, >70 million applications, ~600,000 job transmissions (Rajkumar et al., 2022) | Strong, and it reports no percentage effect size |
| "70 to 85 per cent of jobs are filled through networking" | No credible source located. Not in Granovetter, not in the referral literature, not in any traceable survey | Unsourced |
| Referred hires quit less and are worth more, and look identical on paper | 13% less likely to quit in call centres, ~30% in high tech; discounted profit per worker $2,201 against $1,756; schooling within 0.20 years and cognitive ability within 0.12 SD of non-referred hires; 1.5 million applicant records across nine firms (Burks, Cowgill, Hoffman & Housman, 2015) | Strong |
| Search intensity buys interviews; search quality buys the job you wanted | Active job search correlated .44 with number of interviews, .24 with employment status and .16 with employment quality; intensity .19 with status and .06 with quality (van Hooft et al., 2021) | Strong |
Claims removed from this course, and why
"75 per cent of résumés are rejected by applicant tracking systems." Traced to a 2012 sales claim by a company that closed in 2013. The vendor most associated with the figure states that an ATS rarely rejects a candidate automatically.
"Recruiters spend six seconds on your CV." Thirty recruiters, no peer review, published by a company selling CV services. Peer-reviewed eye-tracking finds 14 to 16 seconds, and finds that looking longer does not mean scoring higher.
"The hiring decision is made in the first seven seconds." No primary source. Contradicted by the only field measurement of decision timing.
"93 per cent of communication is nonverbal." Two 1967 experiments on single words and still photographs, whose author stated the generalisation does not apply outside judgements of liking.
"Men apply at 60 per cent of the qualifications, women at 100 per cent." An unnamed internal company report that no one has produced. The randomised test found 52.1 per cent and 55.7 per cent.
"Women don't ask." The 57 and 7 per cent figures come from a convenience sample of one graduating cohort. The field experiment that tested it found the gap reverses when the advert says the wage is negotiable.
"Work sample tests are the best predictor of job performance at .54." Refuted twice independently, landing on .33 both times.
"Years of experience is what employers screen on, and rightly." They do screen on it. Its corrected validity is .07 with a lower credibility bound below zero.
"Employers have dropped degree requirements." Announcements outpaced practice by roughly two orders of magnitude: not even one hire in 700 last year.
"Tailor everything, always." The one causal estimate shows a real effect that is shrinking as the cost of tailoring falls. Tailoring is worth doing and its value is a scarcity effect, and this course says which parts of the advice are evidenced and which are convention.
Worked exemplars
Four published artefacts, annotated. Three are real documents. The fourth says so.
These are the documents rather than descriptions of them, and they were chosen because somebody published the thing itself.
The Judicial Appointments Commission publishes worked self-assessments with the grade the panel would award and the panel's reasoning, which almost no organisation does, because scores are candidate data. The US Office of Personnel Management publishes the anchored benchmarks its interviewers score against, in the public domain. Abel, Burger and Piraino, and Carranza and colleagues, published the actual reference letter and skills certificate that moved employment in their trials.
The fourth is different and is labelled as such. No employer publishes a real CV alongside the screening decision made on it, so that exemplar is built the other way round: the CV is constructed, and everything said about it is quoted from named practitioners describing their own sorting.
Read them in order. The graded examples tell you what an assessor is reading for. The benchmarks tell you what separates one band from the next. The instruments tell you what verifiable evidence looks like as an object. The screening pile tells you how little time any of it gets.
What is open, and what is not
Level 0 of every programme is free: its lessons, checks and gates need no account, its tools need a free account, and the other levels and the worked exemplars are for early-access members.
Create a free account