STAR is the wrong half of STAR

STAR is two instructions sharing one acronym. The first half, Situation and Task, is the half the evidence supports, because it forces you to name a specific event that actually happened, with a date and a constraint attached to it. The second half, Action and Result, is where candidates write a script, and the script is the failure mode interviewers have names for. Keep the first half as a rule for your thinking. Keep the second half as a shape for your facts, not a text to recite.

The structure is the part that raises validity

The strongest finding in personnel selection is not about charisma or chemistry. It is about standardisation. Campion, Pursell and Brown set out the technique in 1988 with six steps: develop questions from a job analysis, ask every candidate the same questions, anchor the rating scales with worked examples, have a panel record and rate the answers, administer the process identically to all candidates, and document it. Tested on 149 entry-level production hires, it produced inter-rater reliability of r = .88 and predictive validity of r = .34 uncorrected, r = .56 corrected.

The US Office of Personnel Management's practical guide publishes the same mechanism, and summarises the difference in three lines: all candidates are asked the same questions, all candidates are evaluated on the same rating scale, and interviewers do not have to agree in advance on what an acceptable answer is. Its claim about the alternative is blunter. Research consistently indicates that unstructured interviews have little value in predicting job performance, showing low reliability and low to moderate validity.

The size of the gap matters, and the number depends on the correction applied. Schmidt and Hunter report structured interviews at r = .51 against .38 for unstructured, while other frameworks report uncorrected values near .25 for the same construct, a different correction rather than a contradiction. Their 2016 update, a working paper whose assumptions are disputed, puts structured interviews at .58. A separate meta-analysis of 245 validity coefficients across 86,311 people, whose article link inside that page did not resolve when checked, so the figure itself is not directly verified

So a candidate can predict what a strong answer contains: what you did, why you chose it over the alternative, how hard it was, and what changed. That is what behaviourally anchored rating scales reward, alongside the outcome and any critical failure.

The quiet implication is that the questions and the scale are both public property, because the questions come from the same job analysis that describes the role. Each competency carries behavioural examples for every level of a five-level scale, and the level that counts as a competent answer is the middle one: applying the competency in difficult situations and requiring occasional guidance, not the top. That is the argument behind reading interview questions off the job posting.

Situation and Task is the load-bearing half of the acronym for a reason written into the federal guidance. When subject matter experts write behavioural questions, the guide tells them to use superlative adjectives, most, last, worst and least, so the candidate focuses on a specific incident, and to ask for the actions a person took or did not take and their impact. That is an instruction to make the question unanswerable by a general life story.

Why the script fails

The evidence for the next part is not research, and that is worth saying before the claim is made.

The widely reported failure mode is scripted delivery, and the reports are practitioner complaint threads on Reddit: a recruiter telling a candidate their answers sounded scripted and over-rehearsed, interviewers probing rehearsed answers because scripts collapse under follow-up, candidates forcing one prepared story into questions it does not answer, and STAR itself described as rigid and uncanny. That is a forum, not a sample. It shows the complaint circulates among people on both sides of the table. It does not measure how often it happens, and no study in this literature has produced a rate. Any figure claiming a percentage of interviews lost to scripting is not from the research literature.

A script has one path through a question. An interview is a conversation with follow-up questions inside it, and the federal guide devotes a whole development step to them: step five is "Create Interview Probes", with the instruction that the general meaning of a probe should not change between candidates, which argues for consistency, not against probing. Its sample probes for a single behavioural question are the shape of the follow-up you will actually meet.

• Situation: what factors led up to it, could you or anyone else have prevented it, what did you determine as the most critical issue to address.
• Action: how did you respond, what was the most important factor you considered in taking action, what is the first thing you did.
• Outcome: what was the outcome, is there anything you would have said and/or done differently, were there any benefits from the situation.

Three of those nine break a script. "What is the first thing you did" is unanswerable by a rehearsed summary. "Is there anything you would have said and/or done differently" asks for the one thing a prepared answer never contains. And "what was the most important factor you considered in taking action" is asking for the trade-off, directly, with nowhere to go.

The guide is also blunt about what an interviewer should not be doing. Its list of common interviewing mistakes ends by telling interviewers not to base the evaluation on how the candidate acts during the interview, but on the behaviour the candidate describes. A recited answer is a performance artefact, and a scale anchored to described behaviour has nothing to attach it to. Two honest limits: the guide never names scripted answers as a rating error, and no research has measured what one costs. What the guide does is refuse to score delivery.

The incident brief

Prepare incidents, not scripts. Most people quote a length for an answer. Treat that number as folklore: no study has established an ideal answer length, and nobody quoting one can say where it came from. What matters is the allocation inside whatever window you use.

The artefact is a one-page brief per incident, written in facts rather than sentences. Write three tonight, for three things that actually happened to you.

Four rules keep it a brief rather than a script.

• Facts cold, wording loose. Write nouns and numbers. If a line can be delivered without thinking, it is a script line, and it should be deleted, because the fixed wording is the part that reads as recited.
• One number you would be embarrassed not to know. The count, the deadline, the percentage, the headcount. The federal guidance for writing questions asks for the actions a person "took or did not take", which is a reminder that an omission counts as evidence. A story with no mistake in it and no number in it is a press release.
• A trade-off sentence. This is the highest-value line in the answer, and the one most candidates never write. Naming the option you rejected is the only clean way to show that you compared options rather than executed instructions, and the scales reward quality of reasoning and degree of difficulty. "The faster fix broke month-end, so I rejected it" is worth more than any paragraph about how well you communicated.
• A regret line, because one of the nine federal probes asks for exactly that, and a prepared answer has none.

Three incidents, not thirty. Questions are written from competencies rather than topics, so one incident can be re-aimed at a different competency by changing which part of it you lead with. That is an inference from how the questions are built rather than a documented finding. Take the requirement sentence in the posting that worries you, name the competency underneath it, and check that one of your three incidents can carry it.

Two more things. First, end on the result and stop talking. The turn ends when the outcome is stated, not when you run out of things to add, and everything after the number is material to probe rather than evidence to record.

Second, do not force one story into every question. Inflating a thin incident until it answers a question it does not answer is the same failure as scripting, one step earlier, and it is the other complaint the practitioner threads describe. If nothing you have lived fits, say what judgement you would bring to it and why, and let the follow-up work on that.

The honest limit: confidence is not competence

Rehearsal does something measurable, and it is not what candidates assume. Interview training reliably reduces anxiety and raises self-efficacy: one mock-interview study with 15 participants lowered salivary cortisol and marginally raised confidence. Feedback has a small to moderate average effect of d = .48 across 435 studies, 994 effects and more than 61,000 participants, with heterogeneity high enough that what kind of feedback matters more than whether feedback exists.

Then the part that usually gets left out. Interventions have been shown to reduce anxiety and raise self-efficacy without improving evaluator-rated performance, and an early five-arm trial found one-time mock interviews were not clearly superior to didactic instruction. Practice and training opportunities explain roughly 14% of the variance in performance overall and about 1% among elite performers, so the returns thin out exactly where people believe they most need them.

So the defensible claim about rehearsed answers is narrow, and narrower claims are the useful ones. Scored feedback tells you which specific thing to change next. A mirror does not. Confidence is not competence, and the thing worth practising is a specific incident with a specific number in it, not the feeling of having practised.

That holds across the rest of the process. The conversation with the largest documented effect on your income is the one almost nobody rehearses, which is why what happens when you negotiate a salary is worth reading before you need it rather than during. Practise the thing that is hard.