A résumé has always been a claim someone makes about themselves. It was never evidence. For a long time that didn't matter much, because writing a good one took real effort, and effort was itself a weak but usable filter. That filter is gone. Any candidate can now produce a flawless, keyword-optimized, achievement-quantified résumé in ninety seconds, and a meaningful share of the résumés landing in your inbox were written by a model, not a person. Companies that keep screening on the claim, rather than the evidence, are optimizing their funnel for the applicants best at prompting, not the ones best at the job.
A résumé was always a claim. AI just made the claim free
Résumé screening survived as long as it did because producing a convincing one took real time and judgment, which correlated, loosely, with the judgment the job required. That correlation is broken now. A candidate can generate a technically fluent, achievement-quantified, keyword-matched résumé for any role in under two minutes, and increasingly does, whether or not their actual experience supports it. The people running your applicant tracking system today are effectively grading essays that a language model helped write, and grading them on the exact criteria, keyword density, phrasing, structure, that a model is best at gaming. You are not filtering for competence. You are filtering for whoever used the best prompt.
What proof actually looks like
Proof isn't one thing, it's a category of signal that has a real cost to fake, which is the entire point. A shipped open-source project, reviewed in front of the candidate, tells you how they actually structure code and handle edge cases. A paid work-sample test, built around a real (or realistically scoped-down) problem from your own domain, tells you how they think when the answer isn't already indexed somewhere. A structured technical eval, scored the same way across every candidate, tells you where they sit relative to a bar, not relative to how well they interviewed that day. None of these are new ideas. What's new is how urgently they need to replace résumé screening as the first filter, not the fourth.
- Portfolio or open-source review: real code or real published work, reviewed live, questions asked about specific decisions the candidate made.
- Paid work-sample tests: a scoped, realistic problem, compensated, evaluated against a rubric you'd defend to the candidate afterward.
- Structured technical evals: the same problem set and scoring criteria for every candidate in a role, so results are comparable, not vibes-based.
- Reference checks with specific behavioral questions: not 'would you recommend them' but 'tell me about a time this person disagreed with a decision and what they did about it.'
The candidates résumé screening filters out are often your best ones
The strongest engineers we've placed are frequently indifferent, sometimes bad, résumé writers. They'd rather point you at something they built than describe it in bullet points, and many of them have simply stopped bothering to keyword-optimize a document for an applicant tracking system they've come to distrust. Meanwhile the candidates most fluent at producing an AI-polished, achievement-inflated résumé are not reliably the strongest engineers, they're the ones most willing to invest in the document instead of the work. A screening step that rewards document fluency over work quality is adversely selecting against exactly the people you're trying to hire.
Redesigning the loop without exhausting candidates
The fix isn't a longer process, it's a resequenced one. Put one well-chosen piece of evidence first, a portfolio review, a short paid work sample, before any resume-based screening happens at all, and let that evidence do the work five conversational rounds used to do badly. A single 90-minute paid work sample, reviewed by someone who actually does the job, produces more signal than three separate interview rounds asking a candidate to describe their experience. Candidates respect this too: strong candidates consistently tell us they'd rather spend ninety focused minutes proving they can do the work than four hours across five calls being asked to summarize it.
| Step | Résumé-first loop | Proof-first loop |
|---|---|---|
| First filter | Keyword and phrasing match on a self-written document | Reviewed work sample, portfolio, or short paid test |
| What it measures | How well the candidate can describe themselves | How the candidate actually performs on real-shaped work |
| Gameable by AI tools | Yes, almost entirely | Much less; work still has to hold up under review |
| Candidate time cost | Low upfront, high later across many rounds | Front-loaded, but often replaces two or more later rounds |
| Signal comparability across candidates | Weak, depends on writing skill | Strong, if scored against a fixed rubric |
Where a résumé still earns a place in the process
None of this means résumés are useless. They're a reasonable way to confirm basic facts, employment history, dates, domain exposure, and a fine tiebreaker once proof has already separated a shortlist. The mistake is using the résumé as the first and heaviest filter, the gate that decides who even gets a chance to show proof. Flip the order: let proof qualify the shortlist, and let the résumé fill in context around candidates who've already demonstrated they can do the work.
