AI Guides › Workbench
By Nigel Guy · 8 min read
Most people paste their CV into ChatGPT and type "make this better". You get back something longer, smoother and full of words like "spearheaded", and it feels like progress because every line has changed. But nobody has checked what a tired reader actually sees in the first few seconds, so the polish lands in places nobody reads and the real problem, usually the top third of page one, survives untouched.
The rule: diagnose the skim first, rewrite second, and only fix what the skim says is broken.
The "six-second" figure comes from an eye-tracking study by the US job site Ladders, first published in 2012. Its 2018 update put the average initial screen at 7.4 seconds and found that reviewers looked first at current title and employer, then the previous role, then dates, then education. Treat that as one company's study, not a law of nature. The useful part is the pattern: a first-pass reader samples a handful of spots on the page and decides whether to keep going.
That is why the order matters. If you rewrite first, you have no idea which lines the skim lands on, so you optimise everything equally. If you scan first, you learn where the reader quits, and the rewrite becomes a targeted repair job.
| Step | What it does | Cost at time of writing | Best for | The catch |
|---|---|---|---|---|
| 1. Skim simulation | Role-plays a time-poor reviewer and reports where attention drops | Works on ChatGPT's free plan (£0) | Finding the dead zone on page one | It is a simulation; a model cannot see what a real recruiter's eyes do |
| 2. Result-first rewrite | Rebuilds experience bullets around outcome, measure, method | Free plan | Turning duty lists into evidence | Will invent numbers if you let it |
| 3. Keyword gap check | Compares your CV to a pasted job advert and lists missing terms | Free plan | Matching the language of a specific role | Any "ATS score" it gives is made up |
| 4. Cost-of-missing framing | Tightens the summary so the reader sees what they lose by passing | Free plan | Your opening profile and cover note | Easy to tip into arrogance; use lightly |
Paid ChatGPT plans give higher limits and stronger models, but this chain is four messages long and runs fine on the free tier. OpenAI's pricing page did not show fixed UK prices to us at time of writing, so check it in your own region before upgrading for this.
Run all four in the same conversation so each step can see the last one. Before you start, either open a Temporary Chat (not saved to history or used for training while it stays temporary) or turn off Settings > Data controls > Improve the model for everyone. A CV is a document full of personal data.
The point is a timeline, not a review. You want to know what got read, in what order, and the exact line where interest died.
You are a hiring manager for [TARGET_ROLE] at [TYPE_OF_EMPLOYER]. It is late afternoon and you have a long stack of applications left. You give each one a very short first pass before deciding "keep" or "bin".
My CV is pasted below.
Do this in order:
1. List, in sequence, the 5-8 places your eye lands in that first pass (e.g. "current job title", "dates of last role").
2. For each, write one line on what you took from it.
3. Name the single line or section where you stopped reading, and quote it.
4. Give your verdict: keep or bin, plus the one reason that decided it.
5. List the three changes that would most likely flip a "bin" to a "keep", most important first.
Rules: be blunt but specific; quote my text rather than paraphrasing it. Do not rewrite anything yet. If the target role or employer type is missing, ask me before answering. Before you reply, check that every point refers to something actually on my CV.
[PASTE_CV_TEXT]
Fill in: the job title you are aiming for, the kind of employer, and your CV as plain text.
Paste text rather than uploading a heavily designed PDF. Columns, text boxes and icons can come through scrambled, and then you are diagnosing an extraction problem, not your CV.
This uses the formula Laszlo Bock, Google's former head of people operations, published on LinkedIn in 2014: state the accomplishment, how it was measured, and what you did to get it. It is a sound structure. The trap is that a model asked for measurable results will happily supply measurements you never gave it.
Using what you found in the skim above, rewrite only the experience section of my CV.
Goal: every bullet should lead with what changed because of my work, then how it was measured, then what I did. Structure: "Achieved [outcome], measured by [metric], by [action]."
Steps:
1. Start with the section you said lost your attention.
2. For each bullet, keep my facts. Where I have not given a number, write [NUMBER?] and do not estimate one.
3. Where a bullet is pure duty with no visible outcome, flag it and ask me one question that would surface the result.
4. Keep each bullet to two lines at most and use British spelling.
Output: a table with three columns, "original", "rewrite", "question for you (if any)".
Before answering, check that no figure, percentage or client name appears in the rewrite that was not in my original text.
Fill in: nothing new; it builds on Step 1. Answer its questions, then ask it to merge your answers.
The source version of this asks the model to act as an applicant tracking system and score you out of 100. Skip the score. ChatGPT has no access to the employer's system, its settings or the other candidates, so any number it gives is invented and will change if you ask twice. The part worth keeping is the comparison against a real job advert.
Act as a careful recruiter screening for the job advert pasted below.
Compare it against my current CV (the latest version in this chat).
Return:
1. A table of skills, tools, qualifications and phrases the advert asks for, with a column showing "present", "present but worded differently" or "missing" on my CV.
2. For each "worded differently" item, the advert's wording I could adopt, only where it is true of me.
3. For each "missing" item, a question asking whether I have it; do not add it to my CV.
4. Any formatting likely to confuse automated parsing (tables, headers in images, unusual section names).
Do not give a percentage or score. If no job advert is pasted, ask for one.
[PASTE_JOB_ADVERT]
Fill in: the full text of one specific job advert. One advert per run; a generic "make it ATS-friendly" request has nothing to match against.
The original idea leans on loss aversion: make the employer feel what they lose by passing. A light version works in a profile statement or cover note. Applied to the whole CV, it reads as pushy.
Rewrite only the opening profile (3-4 lines) at the top of my CV.
Goal: a reader for [TARGET_ROLE] should finish it knowing the specific problem I solve for an employer like theirs and the evidence that I have solved it before, so that passing on me feels like leaving that problem unsolved.
Constraints: use only facts already in this chat; no superlatives ("exceptional", "world-class"); no claims about other candidates; first person implied, no "I". Give me two versions: one plain, one slightly bolder. Then tell me in one line which you would send and why.
If you do not know the employer type, ask before writing.
Fill in: the target role. Use the plain version unless the sector rewards confidence.
| Your situation | Run |
|---|---|
| Applying broadly, getting no replies | Steps 1 and 2 |
| Strong CV, one specific job you want | Step 3 with that advert, then Step 4 |
| Career change | Step 1 first; the skim usually shows the old job title dominating |
| Short on time | Step 1 only, then fix its top change by hand |