10 min read
Ask a chatbot to "write me a resume" and you'll get something fluent, confident, and completely interchangeable with the two hundred other files in the pile.
Knowing how to use ChatGPT and Claude to write a resume properly means using them for what they're genuinely good at, which is rewriting and tailoring, and refusing to let them do the part they're bad at, which is knowing anything about your career. You bring the facts. The model brings the phrasing. That division is the whole method.
And whatever you draft, test it before you send it. The free ATS resume checker shows what a parser extracts, which is the only opinion that matters before a human reads a word.
Quick Wins
- Before opening any AI tool, write ten real facts: tasks, numbers, tools, outcomes.
- Add "do not invent any numbers" to every prompt you use.
- Delete every adjective the model adds. All of them. See what survives.
What AI is genuinely good and bad at here
Being clear about this saves you from the two failure modes: not using these tools at all, and using them for the wrong job.
| Good at | Bad at |
|---|---|
| Rewriting a weak bullet you supply the facts for | Knowing what you actually did |
| Spotting which posting requirements you haven't addressed | Judging whether a claim is credible for your level |
| Translating internal jargon into market language | Producing specifics, which it will invent if pushed |
| Cutting a long bullet to 20 words | Resisting corporate filler unless told to |
| Drafting fifteen variations so you can pick one | Formatting a file that parses cleanly |
That last row deserves emphasis. A model can give you excellent words inside a layout that fails at the first gate. Content and formatting are separate problems, and the formatting ones kill more applications.
How to use ChatGPT and Claude to write a resume: the workflow
Five steps, and the first one happens before you open anything.
- Collect raw facts offline. For each job, write what you owned, what changed, the number attached, and the tools involved. Ugly notes are fine. This is the material the whole process runs on.
- Rewrite bullets with hard constraints. One job at a time, with an explicit ban on invented figures.
- Tailor against a specific posting. Ask the model to find gaps rather than to fill them, then fill the real ones yourself.
- Edit ruthlessly. This is where you remove the generated tone, and it's the step people skip.
- Check the file. Parsing, coverage, length. Words are only half the job.
Do this now: before anything else, open a blank note and write ten facts about your current job. Numbers where you have them, scale where you don't. Everything below is faster once that exists.
Four prompts you can paste
1. Rewriting bullets
"You are an experienced recruiter. Rewrite the following as three resume bullet options. Rules: lead with the outcome, under 25 words each, name the tools I mention, use plain language, no adjectives like dynamic or innovative, and do not invent any numbers or facts I have not given you. If something important is missing, ask me for it instead of guessing. Here are my raw notes: [paste]"
The last two sentences are what separate this from a generic prompt. Asking the model to request missing information rather than fabricate it changes the output substantially. If you want a reusable library instead of one chat you will lose, copy a ChatGPT or Claude brief from AI Prompts Online's free prompt library , then paste your facts on top. The prompt is not the resume. Your numbers still have to be true.
2. Finding gaps against a posting
"Here is a job description and here is my experience section. List the top eight requirements in the posting, and for each one tell me whether my resume clearly addresses it, addresses it weakly, or not at all. Do not rewrite anything yet. Job description: [paste]. My experience: [paste]"
"Do not rewrite anything yet" matters. Left to itself the model will helpfully invent experience to close the gaps, which is exactly the failure you're trying to avoid.
3. Translating internal titles and jargon
"My internal job title is [title] and my company calls this process [jargon]. Suggest the standard market equivalents a recruiter and an applicant tracking system would recognise, and explain what each one implies about seniority."
4. Pressure-testing your own file
"Read my resume as a sceptical hiring manager for this role. In three sentences, summarise who this candidate appears to be. Then list the three weakest bullets and say exactly what is missing from each."
That summary is the useful part. If the three sentences describe a generic professional rather than you, that's roughly what the screening summary will say too.
The facts only you can supply
The quality gap between a great AI-assisted resume and a terrible one is almost entirely in what you feed it. Five categories, and you need all five per job.
- Scope. What you owned, and how big it was. Headcount, budget, accounts, systems.
- Outcome. What was different afterwards, in the world rather than in your effort.
- Numbers. Real ones. Time saved, revenue, volume, error rate, headcount.
- Tools. Named. Workday, Salesforce, Power BI, Python. Parsers and recruiters both search for these.
- Context. Industry, company size, and what made it hard.
If your work genuinely produces no numbers, use scale and frequency: how many people, how often, what breaks without you. There's a system for capturing all of this as you go in tracking accomplishments weekly , which pays off for applications as much as for pay conversations.
Editing out the AI smell
Generated resume text has a recognisable texture: grammatically perfect, relentlessly positive, and empty. Recruiters spot it constantly, and no detection tool is involved.
Run these four passes on anything a model produces.
- Delete every adjective. Dynamic, innovative, robust, comprehensive, seamless, cutting-edge. Read what's left. If the bullet is now empty, it was always empty.
- Restore the specifics. Models generalise. Put the tool names, the numbers, and the actual system back in.
- Cut the throat-clearing verbs. Spearheaded, leveraged, utilised, facilitated. Led, built, cut, ran, fixed.
- Read every line aloud and ask: could I talk about this for two minutes? If not, delete it. That question is the entire ethics test and the entire interview test at once.
| What the model gave you | After the edit pass |
|---|---|
| Spearheaded innovative cross-functional initiatives driving significant operational efficiencies. | Led the invoice approval redesign across finance, ops, and IT; cut approval time from 9 days to 3. |
| Leveraged data-driven insights to enhance stakeholder decision-making. | Built the weekly margin dashboard in Power BI that the commercial team now runs pricing calls from. |
What not to paste
Worth thirty seconds of thought before you paste anything into any tool.
- Your contact details. The model doesn't need your address or phone number to rewrite a bullet.
- Named clients under NDA. Anonymise by sector and scale instead: "a mid-size regional insurer."
- Confidential figures. Unpublished revenue, headcount plans, internal pricing. Use percentages or ranges.
- Anything from a work account. Check your employer's policy before using a company tool for a job search, for obvious reasons.
- Other people's information. Colleagues' names and performance details aren't yours to share.
None of this stops you using these tools well. Anonymised facts produce the same quality of rewrite, because the model is working on the structure of the sentence rather than the identity of the client.
Edge cases
Career changers
This is where these tools earn their keep. Ask the model to map your existing responsibilities onto the target field's vocabulary, then verify each translation is honest. A teacher's "differentiated instruction across 30 students" really is stakeholder management, but you have to be able to defend the framing.
Very senior roles
Generated text flattens seniority, because it defaults to a mid-level register. Executive resumes need scope, P&L, and organisational scale in the first two lines. Give the model those explicitly or it will write you down a level.
Non-native English speakers
Excellent use case, with one caution: ask for plain professional English rather than polished business language. Over-formal phrasing is its own kind of tell, and simpler sentences parse and read better anyway.
Employment gaps
Don't ask a model to disguise one. Ask it to phrase the real explanation concisely. Fabricated continuity falls apart at reference stage and it is a much bigger problem than a gap ever was.
Cover letters
Same rules, higher risk, because generic cover letters are even more obvious than generic resumes. If you'd rather not manage the prompting, the cover letter generator builds one against the specific posting.
Mistakes that get you filtered
- "Write me a resume for a marketing manager." You'll get the average of every marketing resume ever written, which is precisely what you're competing against.
- Accepting invented numbers. Models produce plausible figures when asked for impact. Check every single one against reality.
- Skipping the edit pass. The adjectives are the tell. Removing them takes four minutes.
- Letting it choose the format. It will happily suggest a layout that doesn't parse.
- Claiming skills to close a gap. Closing gaps on paper creates them in the interview.
- Using the same generated file for every application. The tailoring step is the one that actually raises your match score.
Check the draft before you send it
A model can't tell you whether your file parses, and that's the gate that rejects most applications silently.
Upload the finished draft to the free ATS resume checker and read the extraction. Every job present? Dates attached? Skills where they should be? Those three answers tell you whether the writing even reached the stage where it matters.
Then check coverage with the job match score against the posting you're targeting, and if the layout is fighting you, rebuild on a clean single-column base with the free resume builder .
The short version
- How to use ChatGPT and Claude to write a resume: you bring the facts, the model brings the phrasing, you verify everything.
- Ban invented numbers in every prompt, and ask it to request missing detail rather than guess.
- Then delete every adjective. What survives is your actual resume.
Do this today: write ten real facts about your current role before you open any tool. That list is worth more than any prompt on this page.
Then confirm the draft works. Check your resume for free and see what the parser makes of it.
Read more
- How to beat AI resume screeners — the gates your draft has to clear.
- ChatGPT resume mistakes — the failures to avoid in more detail.
- 5 formatting mistakes that kill your resume — the half no model can fix for you.
Frequently asked questions
Often, though not from detection software. They notice it from the writing: fluent, confident, and completely non-specific. A resume full of impact and cross-functional collaboration with no numbers, tools, or scale reads as generated whether or not it was.
Yes, as a drafting and editing assistant. You supply the facts, it improves the phrasing, and you verify every claim. What is not acceptable is inventing experience, and it fails immediately at interview because you cannot discuss work you did not do.
One that gives the model raw facts and hard constraints: the real task, the real outcome, the tools used, plus instructions to keep it under 25 words, lead with the outcome, and never invent numbers. Vague prompts produce vague bullets.
Strip your contact details, and remove client names, confidential figures, or anything covered by an NDA first. Paste the experience content you want rewritten rather than the whole document, and check your employer's policy if you are using a work account.
Passing depends far more on formatting and specificity than on who wrote the words. Generated text that lacks named tools, numbers, and the posting's actual terminology tends to score poorly, because there is nothing concrete for a matcher to latch onto.