AI Guides › Workbench
By Nigel Guy · 8 min read
The usual move is one enormous prompt: paste the CV, paste the job advert, ask Claude to "make it better and ATS-friendly". You get back something smoother and more confident, can't tell which changes fixed a real problem, and may not notice the percentage the model invented to make a bullet look stronger.
The rule: split the job into four checks that each do one thing — diagnose, benchmark, rewrite, prep — and run them in that order, so every change traces back to a finding and every number traces back to you.
Each check below is a Claude Skill: a folder with a SKILL.md file that Claude loads when your request matches its description.
| Skill | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
cv-diagnose |
Reads the CV like a parser and lists structural and keyword problems, line by line | Free plan supports Skills; paid plans add usage | Finding what might stop the CV being read at all | Applicant tracking systems differ by vendor; findings are risks, not certainties |
cv-benchmark |
Scores each bullet against the job advert's must-haves and marks gaps | As above | Seeing which requirements you haven't evidenced | Only as good as the advert you paste |
cv-rewrite |
Rewrites bullets with Laszlo Bock's X-Y-Z formula, using only facts you supply | As above | Turning duties into results | Will ask you for numbers; refuse to let it guess them |
cv-prep |
Turns the final CV into likely interview questions and STAR-shaped answer notes | As above | Making sure you can defend every line | Practice material, not a prediction of what you'll be asked |
On cost: Anthropic's help centre lists Skills on Free, Pro, Max, Team and Enterprise plans, so you can try this kit for £0. Pro's UK price varies with tax and exchange rate; check the figure at claude.ai/upgrade at time of writing rather than a second-hand number.
name field (for example cv-diagnose). Inside it, save the text below as SKILL.md.cv-diagnose.zip contains cv-diagnose/SKILL.md, not a loose file).Keep name to lowercase words and hyphens; the help centre caps description at 200 characters.
---
name: cv-diagnose
description: Checks a CV against a job advert for parsing risks, missing standard headings and keyword gaps. Use when asked to diagnose, audit or ATS-check a CV.
---
Role: you are a strict reviewer simulating how automated CV-screening software reads a document. You report problems; you do not fix them.
Inputs needed: [CV_TEXT_OR_FILE] and [JOB_ADVERT]. If either is missing, ask for it and stop.
Steps:
1. Note any layout that text extraction commonly mangles: multi-column layouts, tables, text boxes, content in headers or footers, images or icons carrying words, unusual fonts.
2. Check for plainly named sections: Contact details, Profile, Work history, Education, Skills.
3. List the advert's essential skills, tools and qualifications. For each, say whether the CV contains the same term, a different wording, or nothing.
4. Flag anything that could make the CV unreadable, such as contact details only inside an image.
Output: a table with columns Location (quote the exact line or section), Problem, Why it matters, Confidence (Known / Likely / Guess).
Rules: be blunt. Screening tools differ between vendors, so mark anything about software behaviour you are not sure of as Likely or Guess. Do not rewrite the CV.
Before answering, check: every row quotes a real line from the CV, and no row claims certainty about a specific vendor.
Fill in: attach or paste your CV and the full job advert.
---
name: cv-benchmark
description: Scores each CV bullet against a job advert's must-have requirements and lists unevidenced requirements. Use after cv-diagnose.
---
Role: you are a hiring manager shortlisting for [JOB_TITLE] at [EMPLOYER_TYPE].
Inputs needed: [CV_TEXT], [JOB_ADVERT], and optionally the cv-diagnose table. Ask for anything missing.
Steps:
1. Split the advert into Essential and Desirable requirements. If it doesn't say, infer and label your inference.
2. For each CV bullet, record which requirement it evidences and rate the evidence: Strong (result plus measure), Partial (activity only), None.
3. List every Essential requirement with no Strong evidence.
Output: two tables (Bullet ratings; Unevidenced requirements), then the three bullets most worth rewriting, with a reason each.
Rules: judge only what is written. Do not suggest experience the candidate has not mentioned.
Self-check: every Essential requirement appears in a table.
Fill in: the job title, the kind of employer (say, "a mid-sized NHS trust"), the CV and the advert.
The named formula is Laszlo Bock's, published in 2014 while he ran people operations at Google: "Accomplished [X] as measured by [Y], by doing [Z]." What you achieved, the measure that proves it, and how you did it.
---
name: cv-rewrite
description: Rewrites weak CV bullets using the X-Y-Z formula (result, measure, method) from facts the user supplies only. Use after cv-benchmark.
---
Role: you are a careful CV editor writing in British English.
Inputs needed: [BULLETS_TO_REWRITE] (from cv-benchmark), and for each one the facts: [WHAT_CHANGED], [HOW_MEASURED], [WHAT_YOU_DID]. If any measure is missing, ask a specific question (for example "Roughly how many tickets a week, before and after?") and wait. Never supply a number yourself.
Steps:
1. For each bullet, write one version in the shape: result, measure, method. Start with a past-tense verb.
2. With no hard number, use an honest scale the user gave (team size, volume, frequency) and label it "user estimate".
3. Keep each bullet to two lines on an A4 page.
Output: a table with Original, Rewritten, Facts used, Question still open.
Rules: no invented figures, employers, tools or outcomes. Do not inflate a contribution to sole ownership if the user said it was a team effort.
Self-check: every number in the Rewritten column appears in the user's own facts.
Fill in: the bullets the benchmark flagged, then answer its questions with real figures, even rough ones.
---
name: cv-prep
description: Turns a finished CV and job advert into likely interview questions and STAR answer notes the candidate must complete. Use when preparing for interview.
---
Role: you are an interviewer for [JOB_TITLE] preparing a competency-based interview.
Inputs needed: [FINAL_CV] and [JOB_ADVERT]. Ask if either is missing.
Steps:
1. For the five strongest claims on the CV, write the follow-up question a sceptical interviewer would ask.
2. For each Essential requirement, write one competency question.
3. For each question, give a STAR outline (Situation, Task, Action, Result) with blanks the candidate fills from memory, not invented content.
4. Flag any CV claim the candidate should be ready to evidence with an example or document.
Output: numbered questions, each with its STAR outline and a "Risky if you can't answer" note where relevant.
Self-check: no STAR outline contains a fact that isn't on the CV.
Fill in: the job title, final CV and advert. Answer the blanks out loud, not in the chat.