Claudeを使って履歴書を選考する方法：Alex Skillガイド

Generally yes, but several jurisdictions attach conditions and the obligations sit with the employer rather than the vendor. New York City has required a bias audit, a public summary and 10 business days' notice to candidates since July 2023. Illinois has required notice since January 2026 and prohibits using zip code as a proxy for a protected class. California's FEHA regulations, in force since October 2025, make evidence of anti-bias testing relevant to a discrimination claim and require four years of records. Colorado's SB 26-189 takes effect in January 2027. In the EU, recruitment counts as high-risk under the AI Act, with those duties now due on 2 December 2027. None of this is legal advice; take proper advice for your jurisdiction.

It depends entirely on how you use the output. New York's rules describe three situations that count: you rely only on the tool's simplified output with nothing else considered; the output is one of several criteria but carries more weight than the others; or the output overrules a conclusion reached another way, including by a person. Summarising a CV for a recruiter who reads every application and weighs that summary no more heavily than anything else is different from ranking 200 candidates and interviewing the top ten. The second is what triggers the audit and notice duties.

No, and this is the detail most articles still have wrong. The high-risk obligations covering recruitment were originally due on 2 August 2026, but the Digital Omnibus, which entered into force on 27 July 2026, moved them to 2 December 2027. Other parts of the AI Act, including transparency duties, were not postponed, so the later date is not a reason to stop paying attention.

Write the rubric before you read a single CV, and keep it. Separate genuine must-haves from preferences, say what evidence would satisfy each one, and instruct the model to judge on job-relevant evidence only, ignoring name, age, address, photo, nationality, school prestige and career gaps unless the brief makes them relevant. Make red flags into questions for a screen call rather than automatic rejections. Then read the applications yourself at the boundary, and record where you disagreed with the model. Bias is not removed by a prompt; it is constrained by having a written standard that existed before the candidates did.

$29, one time, with no subscription or per-seat fee. It runs in Claude, ChatGPT, Gemini or Copilot. It is a skill file that shapes how the model works, not a hosted screening system, so it does not come with a bias audit and it does not discharge any obligation you have as an employer.

A written scoring rubric with must-haves separated from nice-to-haves, advance / hold / no verdicts with two or three specific reasons each, and green and red flag analysis framed as screen-call questions. Beyond the screen itself it also builds Boolean search strings for LinkedIn, GitHub, Dribbble and Google X-Ray with each component explained, competitor talent mapping and sourcing hit lists, and role-specific phone screen guides with conversion analysis by channel.

No. Every date, title, qualification and achievement is taken at face value, because the model has no way to check any of it. Verification stays a human job: references, qualification checks, and the questions you ask on the screen call. What the model can do is notice that a claim is vague or that a progression looks unusual, and tell you to ask.

In several places you must, and where you do not have to it is still the better default. New York City requires notice at least 10 business days before use of a covered tool, Illinois requires notice, and Colorado will require both advance notice and a plain-language explanation within 30 days of an adverse outcome from January 2027. Beyond compliance, a candidate who learns after the fact that a model read their application reacts far worse than one who was told upfront.