Cover the shift: 10 AI skills for hiring without the panic
cv-sift
sixty applications down to a shortlist you can defend
How the two work together
Claude thinks it through. Paste the Claude prompt into Claude Code, or drop the folder into your skills folder. Claude does the judgement: what to look for, what is worth doing, what is right.
Codex gets it done. At the hand-off point Claude runs Codex on your machine with one command and passes it the Codex prompt. Codex does the mechanical part and hands the result back. Claude checks it before you see it.
No API key to set up: Claude calls the Codex you already have installed. If Codex is not installed, Claude does that half itself and tells you.
Prompt for Claude
--- name: cv-sift description: Turns a pile of job applications into a scored shortlist judged against criteria written down before any application is opened. Use when you have more applications than you can read carefully and you need a written reason for every yes and every no. --- # A shortlist you can defend, from sixty applications in one evening You give me the job advert you posted and the applications you have received, in whatever state they arrived: PDFs, phone photos of a handwritten form, emails with a CV pasted into the body. You get back a scoring grid written before anything was read, every application scored on the same criteria with the evidence quoted beside each score, a ranked shortlist cut at the number of interviews you can actually run, a reserve list, and a one page record of who decided what and when. ## What it does 1. **Write the scoring criteria before you open a single application.** Take the advert you actually published and pull out four to six essentials and no more than four desirables. An essential is something the person could not do the job without in their first week, and that you can see from an application form: available Friday and Saturday evenings, has worked services above eighty covers, holds Level 2 Food Hygiene. A desirable is something you would pay a little more for but would train: cellar work, a second language your regulars speak, barista experience. The Information Commissioner's Office guidance on shortlisting says "You should inform candidates about the selection criteria you will apply to their information in order to shortlist them. You should provide these details before you collect the information", so if a criterion was never in the advert, either drop it or accept that you are adding it late and note that on the record. 2. **Turn every criterion into a 0, 1, 2 scale and write the three sentences out.** For "volume experience", 0 might be "no service figures given anywhere", 1 "worked in a kitchen or on a floor but no size stated", 2 "states covers, sittings or transactions at or above our Friday night". Do this for all of them before you read anything, because once you have a face in your head you will bend the wording to fit. Essentials use 0, 1, 2. Desirables use 0 or 1 only, so a nice-to-have can never outweigh a must-have. Save this grid with the date and time on it. 3. **Number the applications and take out everything that is not about the job.** Give each one a number, then remove or cover the name, photo, date of birth, school and university leaving dates, home address, nationality, marital status and anything about children or care. Keep the contact details in a separate list keyed to the number, which only you see. The Information Commissioner's Office puts it directly: "you should have a designated member of staff remove any irrelevant information from the application form before it's presented to the decision-makers", and it says of pseudonymising that it "can be an effective way of avoiding bias in decision-making and help ensure that decision-makers don't take irrelevant information into account". 4. **Run one knock-out pass at thirty seconds an application, on stated essentials only.** Go through all sixty asking one question: does this application show, or fail to show, the two or three things the advert said were essential. Sixty at thirty seconds is half an hour and will usually leave you thirty to forty. Two guards on this pass. If a single essential knocks out more than half the pile, stop, because the criterion is probably wrong or the advert asked for the wrong thing. And right to work is not a sift criterion: that check happens on the person you are about to employ, never guessed from a name, a place of birth or a gap in a school history. 5. **Score one criterion at a time across every application, not one application at a time.** Open criterion one, score all thirty-five, close it, then criterion two. This keeps the standard steady, stops a strong first paragraph inflating every later score, and is faster: twenty seconds per application per criterion, so thirty-five applications across six criteria is about seventy minutes. Do not total anything until every column is filled, and do not look back at earlier columns while you score a later one. 6. **Put the evidence next to the score, or the score is zero.** Every score above zero carries a short quote from that application and where it came from, for example `2 - "ran the pass for 120 covers Fri/Sat, The Anchor, 2023-2025", work history`. If you cannot find the words, you are scoring an impression, and the score goes to zero with the note "not evidenced". Nothing gets read into an application: not that a chef in a large hotel must have done volume, not that a long stint somewhere small means low standards. 7. **Treat gaps, short stints and messy histories as questions for interview, never as marks.** Hospitality histories are seasonal contracts, agency shifts, three months at a place that closed, and time out for illness, caring or a baby. Do not score any of it. Write the question in a "to ask" column instead: "two gaps of eight months, ask what they were doing". A gap that has been silently marked down is one of the easiest ways to disadvantage disabled applicants, carers and women, and you will never be able to show why that person scored lower. 8. **Rank, cut at the number of interviews you can genuinely run, and re-score the boundary band.** Total the columns and sort. Decide the cut from your diary, not from a score: six interviews at thirty minutes with ten minutes between is four hours, so the shortlist is six and the reserve is the next three. Then take the five applications either side of the line and have a second person score just those, without seeing your scores, and talk through any where you differ by more than one point. Acas says "Ideally 2 or more people should be involved in choosing who to interview. This reduces the risk of personal biases or unintended discrimination"; if you are on your own, the boundary band is the part where a second pair of eyes changes an outcome, so spend the favour there. 9. **Write the sift record, then reply to everybody.** One page: the criteria grid with its timestamp, the numbered table of scores and evidence, who scored, the date, the shortlist, the reserve list, and the questions to carry into interview. Then match the numbers back to names and reply to all sixty, including the ones you are not taking forward. Set a deletion date for the unsuccessful applications and put it on the record. The Information Commissioner's Office example is a restaurant with fifty applications: "Unless there is a clear business reason for doing so, the restaurant should not keep recruitment records for unsuccessful candidates beyond the statutory period in which an applicant can bring a claim arising from the recruitment process." Ask your HR adviser or solicitor what that period is for you and set the date from their answer rather than guessing it. ## Then it checks 1. The criteria grid carries a date and time earlier than the first application opened, and every criterion in it appears in the wording of the advert that was published. 2. Every application that survived the knock-out pass has a score in every criterion column. No blanks, no dashes, no "n/a". 3. Every score above zero has a quote from that application beside it, with the part of the form it came from. Every score of zero says either "not evidenced" or which essential is missing. 4. No name, photo, age, date of birth, address, nationality, school leaving date, or reference to children or caring appears anywhere in the scored grid. 5. The shortlist is exactly as long as the number of interview slots stated, a reserve list is named, and the five applications either side of the cut have two independent sets of scores. 6. The record names who scored, on what date, and the date the unsuccessful applications will be deleted, and a reply has been drafted for every applicant including the unsuccessful ones. Any check fails: name it, redo that step once. Failed twice: say what is wrong and stop. ## Rules - Public information only. - Never invent a fact, a number or a quote. - Never write a score that is not backed by words the applicant themselves wrote. An invented justification is worse than no shortlist at all, because the moment you write "seems reliable" next to a number you have created the document that would be read out against you, and you have lost the only real benefit of scoring, which is that the reason is true. - Never infer age, nationality, immigration status, health, pregnancy, religion or caring responsibilities from a name, a photograph, a date, a gap or a place, and never record such an inference even as a note. A guess of this kind, written down once, converts an ordinary hiring decision into evidence of discrimination. - Never rewrite or reweight the criteria after applications have been read. If a criterion turns out to be wrong, say so on the record, start the scoring again from the top of the pile, and keep both versions of the grid. - This output is a working document prepared for the owner's HR adviser, solicitor or accountant to check before it is used. It does not decide who you may lawfully employ, it does not replace a right to work check, and it is not employment law advice. ## Built from - Acas, "Choosing who to interview", https://www.acas.org.uk/recruitment/choosing-who-to-interview, no publication date shown, read 13 September 2026: the checklist method of turning each point of the job description into a scored item, and the instruction that two or more people should be involved, which shaped the boundary band re-score in step 8. - Information Commissioner's Office, "Employment practices and data protection: recruitment and selection - Shortlisting, testing, and interviewing candidates" (draft guidance, consultation closed), https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/recruitment-and-selection/shortlisting-testing-and-interviewing-candidates/, no publication date shown on the page, read 13 September 2026: telling applicants the criteria before collecting their information, stripping irrelevant information before it reaches the decision maker, and pseudonymising to keep bias out of the decision. - Information Commissioner's Office, "Employment practices and data protection: recruitment and selection - Keeping recruitment records" (draft guidance), https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/recruitment-and-selection/keeping-recruitment-records/, no publication date shown on the page, read 13 September 2026: the fifty-applications restaurant example and the rule that unsuccessful applicants' records should not be kept beyond the period in which a claim could be brought, which is why step 9 ends with a deletion date. - Chartered Institute of Personnel and Development, "Selection methods", https://www.cipd.org/en/knowledge/factsheets/selection-factsheet/, 11 November 2024: the two-stage shape of selection, the case for scoring against objective criteria rather than impression, and the warning about automated keyword sifting that this skill deliberately does not use. - GOV.UK, "Check a job applicant's right to work", https://www.gov.uk/check-job-applicant-right-to-work, no publication date shown on the page, read 13 September 2026: the rule that the check is made on the person before you employ them, which is why right to work never appears as a scoring criterion in step 4.
Prompt for Codex
# cv-sift ## You are given A folder from a UK hospitality business containing the job advert as it was published (PDF, DOCX or text), and the applications as they arrived: CV PDFs, phone photographs of a handwritten application form, emails with a CV pasted into the body, jobs board screenshots, the occasional DOCX. File names are inconsistent, formats are mixed, and some people have applied twice. Claude supplies two further files. Before the sift, `criteria.md`: the scoring grid carrying its own date and time stamp, four to six essentials, no more than four desirables, and the written 0, 1, 2 anchors for each essential and 0, 1 anchors for each desirable. After the sift, `scores.csv`: one score and one evidence quote per application per criterion, plus the number of interview slots the owner can run. You never write, change or infer a score. ## Produce Write into an `output/` folder next to the inputs. Work in two passes. Pass one, run before Claude reads anything: 1. `applications-index.csv` with columns in this order: `Application no`, `Source file`, `Format`, `Date received`, `Duplicate of`. Number from 001 upward in the order the files were received. Where the same person has applied twice, keep the later application and put the earlier number in `Duplicate of`. 2. `private/contact-key.csv` with columns: `Application no`, `Full name`, `Email`, `Phone`. This is the only file anywhere in the output that carries identity. 3. `redacted/application-<nnn>.md`, one per application, numbered to match the index. Transcribe the content and replace with `[REDACTED: <field>]` every instance of name, photograph, date of birth, age, home address, nationality, marital status, school or university name, school or university leaving date, and any reference to children or caring. Keep employers, job titles, employment dates, stated covers or transaction volumes, qualifications, stated availability and the applicant's own wording throughout. 4. `scoring-grid-blank.csv` with columns in this order: `Application no`, then for each criterion named exactly as `criteria.md` names it, a score column followed immediately by an `Evidence: <criterion name>` column, then `To ask at interview`. No totals column. 5. `knock-out-log.csv` with columns: `Application no`, `Essential tested`, `Quote or "absent"`, `Kept or removed`. Pass two, run once `scores.csv` is supplied: 6. `sift-record.md`, one printable page: `criteria.md` reproduced in full with its timestamp, the scored table, the name of each scorer, the date of scoring, the shortlist, the reserve list, the questions carried into interview, and a `Deletion date for unsuccessful applications: [OWNER'S ADVISER TO SET]` line. 7. `shortlist.csv` with columns: `Application no`, `Essentials total`, `Desirables total`, `Overall total`, `Rank`, `List` (`shortlist`, `reserve` or `not taken forward`), `Boundary band` (`yes` or `no`). Sort by `Overall total`, highest first. Cut the shortlist at exactly the number of interview slots stated in `scores.csv` and put the next three on the reserve list. Mark `Boundary band` as `yes` for the five applications either side of the cut. 8. `replies/reply-<nnn>.md`, one draft per application including every unsuccessful one, addressed using the name from `private/contact-key.csv`, with the outcome and no scores or reasons unless Claude supplied them. ## Rules - Never write a score, adjust a score, break a tie or fill an empty score cell. A blank score is a finding: report it and stop. - Never carry a name, photograph, date of birth, age, address, nationality, school leaving date or reference to children or caring out of `private/contact-key.csv` into any other file, including file names and the sift record. - Never infer age, nationality, immigration status, health, pregnancy, religion or caring responsibility from a name, a photograph, a date, a gap or a place, and never write such an inference down even as a note. - Every evidence cell is a quotation taken verbatim from that application with the part of the form it came from. Never paraphrase it and never write one yourself. Where `scores.csv` gives a score above zero with no quote, leave the cell as `NO QUOTE SUPPLIED` and list it in the return. - Never rewrite, reorder or reweight `criteria.md`. If a criterion in `scores.csv` does not appear in `criteria.md`, stop and report it. - Gaps, short stints and broken employment histories go in `To ask at interview` as a question. Never turn one into a score or a comment. - Never set the deletion date yourself. Leave the placeholder for the owner's HR adviser or solicitor. - Use British English, £ and DD Month YYYY dates. No em dashes. - Write no covering commentary, no summary of the candidates and no recommendation. This is a working document prepared for the owner's HR adviser or solicitor to check before it is used. ## Return A list of the files written with their absolute paths, the number of applications indexed, the number of duplicates found, the number of redactions made by field type, the number of score cells left blank, the number of scores above zero with no quote, and the application numbers in the boundary band.
Built from the best public work on this
Sources for cv-sift
Everything below was opened and read on 13 September 2026. Nothing is cited that could not be loaded.
1. Acas, "Choosing who to interview"
https://www.acas.org.uk/recruitment/choosing-who-to-interview, no publication date shown on the page, read 13 September 2026.
Acas is the Advisory, Conciliation and Arbitration Service, the public body whose guidance is what an employment tribunal and most UK HR advisers will expect a small employer to have followed. This page is the closest thing to an official UK method for the sift itself. Two instructions from it are load-bearing in this skill. The first is the mechanism: "Turn each point in the job description and person specification into a checklist" and then "Score each application form against the checklist", which is exactly steps 1, 2 and 5. The second is the staffing: "Ideally 2 or more people should be involved in choosing who to interview. This reduces the risk of personal biases or unintended discrimination." The page also warns against using information you did not tell applicants you would use, singling out looking at "some applicants' social media profiles, but not others", which is why nothing in this skill ever leaves the application form. Where the skill departs from Acas is on the second person. A restaurant owner sifting at eleven at night usually has nobody to sit with, and a rule they cannot follow gets abandoned along with the rest of the method. So the skill asks for a second scorer on the five applications either side of the cut only, which is the band where a second opinion actually changes who gets interviewed. That is weaker than what Acas advises, and the skill says so on the record rather than pretending it is equivalent.
2. Information Commissioner's Office, "Employment practices and data protection: recruitment and selection - Shortlisting, testing, and interviewing candidates"
https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/recruitment-and-selection/shortlisting-testing-and-interviewing-candidates/, no publication date shown on the page, read 13 September 2026. The page carries the notice "Our consultation on this draft guidance is now closed. The final version will be published in due course", so this is draft guidance and the skill treats it as good practice rather than settled rule.
The ICO is the UK data protection regulator, and this page is the only source found that covers the sift as a data handling exercise rather than an HR one. Three lines drive the skill. Step 1 comes from "You should inform candidates about the selection criteria you will apply to their information in order to shortlist them. You should provide these details before you collect the information. For example, by including it in the job description or application form", which is why the skill checks every criterion against the published advert and flags any that was added late. Step 3 comes from "you should have a designated member of staff remove any irrelevant information from the application form before it's presented to the decision-makers", and from "Pseudonymisation can be an effective way of avoiding bias in decision-making and help ensure that decision-makers don't take irrelevant information into account." The skill extends that pseudonymisation advice further than the ICO does: the ICO gives it in relation to tests and assessments, whereas this skill applies it to the CV sift as well, on the grounds that the sift is where a name most obviously changes an outcome (see source 4). The page also says "For most roles, it will usually be appropriate to shortlist candidates based on the information they provide rather than information you collect from other sources (eg references or from public social media profiles)", which matches the Acas warning and is the reason the skill scores nothing that is not in the application itself.
3. Information Commissioner's Office, "Employment practices and data protection: recruitment and selection - Keeping recruitment records"
https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/recruitment-and-selection/keeping-recruitment-records/, no publication date shown on the page, read 13 September 2026. Same draft status as source 2.
This page answers the question a small employer never thinks about until it bites: what happens to the fifty-nine applications you did not take. Its worked example is a hospitality one, which is unusual and useful. "A restaurant receives 50 applications for a job vacancy. Unless there is a clear business reason for doing so, the restaurant should not keep recruitment records for unsuccessful candidates beyond the statutory period in which an applicant can bring a claim arising from the recruitment process." It also makes the opposite point, that destroying too fast is its own problem: "if you destroy records about any of the candidates too quickly, it may be more difficult for you to prove that your end-to-end process is transparent, fair and accountable", and it notes candidates can make a subject access request for what you hold, including interview notes. That pair of pressures is why step 9 of the skill ends with a written deletion date rather than either "keep everything" or "bin it". Where the skill refuses to follow the internet consensus is on the number. Several commercial recruitment blogs state a flat six months for unsuccessful applicants, presented as an Equality Act rule. The ICO page itself gives no such number: it points to the Limitation Act 1980, the Limitation (Northern Ireland) Order 1989 and the Prescription and Limitation (Scotland) Act 1973, says the period "varies depending on the nature of the claim, and may be a number of years", and advises employers to "seek independent legal advice on how long you may need to keep these records for". The skill therefore instructs the owner to get the period from their own adviser and never to assert a figure.
4. Nuffield College, University of Oxford, Centre for Social Investigation (Valentina Di Stasio and Anthony Heath), "New CSI report on ethnic minority job discrimination"
https://www.nuffield.ox.ac.uk/news-events/news/new-csi-report-on-ethnic-minority-job-discrimination/, 21 January 2019.
This is the evidence for why unstructured sifting cannot be trusted, and it is British rather than borrowed from American research. The study sent fictitious applications to nearly 3,200 real job openings across manual and non-manual roles including chefs and shop assistants, holding qualifications and experience constant and varying only the ethnic signal. The headline finding is that "Applicants from minority ethnic backgrounds have to send 60% more applications to get a positive response" than white British applicants with matched CVs, with the sharpest penalties for Black British and Pakistani-heritage applicants and, in Professor Heath's words, "The absence of any real decline in discrimination against black British and people of Pakistani background is a disturbing finding". The mechanism matters more than the number for this skill: Dr Di Stasio's reading is that "employers may simply read no further as soon as they see a Middle East-sounding or African-sounding name." That is a description of the first ten seconds of an unstructured sift, and it is the direct justification for step 3, for scoring numbered applications rather than named people, and for the rule that no protected characteristic may be inferred from a name. This skill does not repeat the study's percentage anywhere in its own output, because the owner does not need a statistic, they need the numbered pile.
5. GOV.UK, "Check a job applicant's right to work"
https://www.gov.uk/check-job-applicant-right-to-work, no publication date shown on the page, read 13 September 2026.
The official statement of when and how right to work is established: "You must check that a job applicant is allowed to work for you in the UK before you employ them", with a civil penalty where an employer has "not carried out a correct right to work check", and three valid routes (an online check using the applicant's share code, original documents checked as "genuine, original and unchanged and belong to the person who has given them to you", or an identity service provider using IDVT). The decision this settles in the skill is a negative one. The check is on the person you are about to employ, at the point of employing them, and by a defined method. It is not a sift criterion, and nothing about it can be established from a CV. So step 4 forbids using right to work as a knock-out, and forbids inferring it from a name, a place of birth or a school history, which is the most common way a well-meaning owner turns a paperwork duty into a discrimination claim. The skill stops there and hands the actual check to the owner and their adviser rather than describing how to perform it.
Best public prompt we found for this job
The closest public artefact is the `resume-screener` skill in the GitHub repository `umairalipathan1980/Resume-Screening-AI-Agent-with-Claude-s-Skills`, read at https://raw.githubusercontent.com/umairalipathan1980/Resume-Screening-AI-Agent-with-Claude-s-Skills/main/.claude/skills/resume-screener/skill.md on 13 September 2026. Its star count is unread: api.github.com rate-limited this machine's IP on every attempt, and the rules for this pack forbid taking the number from the repository page or a search snippet. The line worth copying is from its validation checklist:
☐ **Accuracy**: Scores align with evidence from resume; no invented or assumed information
That single line is the whole difference between a scoring grid that protects you and one that decorates a decision you had already made, and it is why this skill drops any score to zero when no quote can be found. Two things in that file were deliberately not copied. Its "Common Red Flags" list treats "Significant unexplained employment gaps" and "Frequent job changes (< 1 year at multiple positions) without clear progression" as marks against a candidate, which in UK hospitality describes seasonal contracts, agency shifts, students and anyone who has taken time out for illness or caring, and scoring it down is a direct route to disadvantaging disabled applicants, carers and women. This skill turns both into interview questions instead. It also drops the file's "Poor resume quality (significant typos, formatting issues, unclear structure)" flag, because a kitchen porter or a section chef is not being hired to write, and the best applicant in a pile of sixty is often the one who has never written a CV in their life.
Want this running in your business?
I optimise how businesses run — your sales, your visibility, your social media — and build bespoke software where nothing off the shelf fits. The first conversation is free. Work starts from £150 a day.
Foxera