How to Make a Resume With ChatGPT (2026): Step-by-Step

What You Need Before You Start

The single biggest factor in whether this works is what you bring to the first prompt. ChatGPT has no access to your work history, so every gap you leave, it fills.

Put four things in one place first:

Your current resume, or a rough dump. Even a messy list of jobs, dates, and duties. It works dramatically better reacting to material than generating from nothing.

The full job posting — the actual text, not a job title. A resume aimed at “marketing manager” in the abstract is aimed at nothing.

Your real numbers. Team sizes, budgets, percentages you remember, time saved, customer counts. Dig them out of old performance reviews, dashboards, or your own analytics. This is the part that takes effort and the part that decides whether the resume is credible.

Anything you’re proud of that isn’t on the resume yet. The process you fixed, the thing you shipped, the person you trained. Usually the strongest bullets, almost always missing.

Ten minutes here is the difference between editing a draft and rewriting one.

Step 1: Set Up the Conversation

Your first message shouldn’t ask for a resume. It should set the rules:

You are an experienced resume writer who works with candidates in [your field]. I’m going to give you my work history and a job description I’m targeting.

Before you write anything, ask me up to 10 questions about my experience that you need answered in order to write specific, evidence-backed bullets. Focus on outcomes, scale, and numbers.

Critical rule for this entire conversation: never invent a metric, a percentage, a dollar figure, a team size, or an accomplishment. If a bullet needs a number I haven’t given you, write [METRIC NEEDED] instead of estimating.

Two things make this work. The interview step surfaces accomplishments you’d never think to mention — it’s good at asking “how many people used the thing you built?”, and you know the answer. And the placeholder rule turns invisible guessing into a visible checklist.

Answer the questions properly, and say “I don’t know” when you don’t. That’s better than a guess, and it’s what the placeholders are for.

Step 2: Draft One Section at a Time

Don’t ask for the whole resume at once. Full-resume output comes back in a single uniform rhythm — same sentence shape, same length, same verbs — and that’s the pattern recruiters skim past.

Experience bullets, one job at a time:

Using only the details I gave you about my role at [Company], write 5 experience bullets. Lead each with a strong verb, describe what I actually did, and end with the outcome. Vary sentence structure between bullets. Use my numbers exactly as I stated them, and don’t add responsibilities I haven’t described.

That final clause matters more than it looks. Without it, “ran the newsletter” comes back as “owned segmentation, A/B testing, and lifecycle strategy.”

The summary, written last — so it distills the bullets rather than setting a target they get bent toward:

Write a 2-3 sentence professional summary for the [job title] role in the posting I shared. Lead with what that posting emphasizes most. No adjectives like “passionate,” “dynamic,” or “results-driven.”

Skills, extracted rather than generated:

Read the job description again. List the hard skills and tools it names, in the order it emphasizes them. Then mark which ones appear in my work history. Do not add skills I haven’t demonstrated.

That last one is the most reliably useful thing ChatGPT does here, and the least glamorous. Keyword extraction requires no facts about you, which is exactly why it doesn’t fail the way generation does.

Step 3: Tailor to the Job

Ask for analysis before any rewriting:

Compare my resume against this job description. Tell me: (1) which requirements it already demonstrates and where, (2) which requirements my history could support but doesn’t mention, (3) which bullets are least relevant to this role. Don’t rewrite anything yet.

Reading the analysis first keeps you in control, and it usually reveals the gap isn’t missing experience but buried experience — the relevant work is real and sitting in bullet five of your second job where nobody reads it.

Only then ask for changes, and only to the summary, the skills order, and two or three bullets. Titles, dates, employers, and metrics never change between applications. Our tailoring guide covers the full system.

Step 4: Edit Out What It Added

This is the step that decides whether the resume works, and it’s the one people skip. Work through it in this order.

Skills and duties you never claimed. Go bullet by bullet and ask: did I tell it I did this? Segmentation, A/B testing, SEO strategy, stakeholder management, on-site event work — these appear constantly and describe things you may never have done. Cut anything you couldn’t talk about for two minutes under questioning.

Verbs that assert a result. Search for grew, increased, improved, drove, boosted, reduced. Each one claims an outcome. Do you have evidence, or did the sentence arrive with the assumption baked in? Check the verb before you check the number.

Numbers you didn’t supply. Every figure, against what you actually provided.

Seniority drift. “Helped with” becoming “led.” “Contributed to” becoming “owned.” Read your bullets against your real title and scope.

The vocabulary. Spearheaded, leveraged, orchestrated, seamlessly, robust, streamlined, results-driven. Swap for plain verbs — led, used, built, ran, cut, fixed.

Your contact details. Read them character by character. Generated contact blocks are worth double-checking before anything goes out.

A useful last pass is to turn ChatGPT on its own output:

Review this resume as a skeptical recruiter who reads 200 a week. Flag every claim I’d struggle to defend with a specific example, every bullet that could belong to anyone in this field, and every skill that isn’t clearly evidenced.

It’s better at critique than at generation, because critique doesn’t need facts it doesn’t have.

Worth knowing what this pass is actually protecting you against: recruiters aren’t running detection software, and no applicant tracking system checks authorship. What gets noticed is genericness — and separately, claims you can’t back up in conversation. We cover what actually gives an AI resume away in detail.

Step 5: Get It Into a Document

ChatGPT can produce the file for you. If you have it write into canvas, you can download the result directly as a PDF, a Word document, or Markdown. That’s the cleanest route.

If you’re working in a normal chat reply instead, copy the text into your own document rather than pasting the raw reply — chat replies carry formatting characters like asterisks and hash symbols that shouldn’t end up in a resume.

Either way, check the finished document against what applicant tracking systems actually parse. A parser walks your file’s structure and decides which text is a job title, which is an employer, which is a date, and which belongs to which role. Then it hands that structured record to the recruiter — and what gets searched and filtered is the record, not your file. When parsing degrades, your resume doesn’t look broken. It just stops matching, and nobody tells you why.

What to confirm before you submit:

• Single column. Multi-column layouts are the most common parsing failure — the parser reads across the columns and interleaves the text.
• Standard section headings. “Work Experience,” “Education,” “Skills.” Not “Where I’ve Been.”
• No text boxes, tables, headers, or footers. Content inside those elements often isn’t extracted at all.
• Consistent, conventional date formats, so each role can be attached to a timespan.
• Simple bullet characters — hyphens or standard round bullets, not decorative symbols.
• PDF, unless the posting specifically asks for .docx.

Our ATS-friendly resume guide covers these rules in depth.

Common Mistakes

Submitting the first draft. Almost every complaint recruiters have about AI resumes is a complaint about unedited output.

Auditing only the numbers. Invented skills are the more common problem and the harder one to spot, because they read as job duties rather than as claims.

Letting it set your level. It writes everyone as a high-performing senior contributor. If you’re two years in and your bullets claim strategic ownership, that surfaces in the first interview question.

One conversation for thirty applications. Long threads drift and start blending details between jobs. Start fresh each time.

“Make this sound more impressive.” Reliably produces inflation. Ask for more specific instead — specificity is what actually reads as impressive.

Never checking the exported file. Good content in a document nobody can parse fails silently.

What We Found When We Tested This

The advice above isn’t guesswork about the editing step. We ran a controlled test to see what actually ends up in an AI-written resume.

We built a candidate profile containing zero numbers — just titles, dates, and plain descriptions like “ran the company email newsletter” — so anything specific in the output was provably generated. Then we ran two prompts against it: the naive one most people use, and the constrained one in Step 1. Six runs, each in an isolated session.

It invented no statistics at all. Not one percentage, dollar figure, or headcount, in any run — which contradicts the warning most articles on this topic lead with.

It invented capabilities in every run.

One run rewrote “Helped with product launches” as “Serve as marketing lead on product launches” — same fact, promoted a level.

And the anti-fabrication rule only worked halfway. The constrained runs used the [METRIC NEEDED] marker faithfully and estimated nothing. But look at what it wrapped them in, quoted exactly as it came out:

“Grew newsletter open rate from [METRIC NEEDED] to [METRIC NEEDED]”

The number is honestly withheld. The claim that growth happened is invented. Fill that bracket in from your real analytics and you’ve certified a trajectory the model assumed for you — while following the anti-fabrication advice correctly.

That’s why Step 4 tells you to check the verbs before the numbers.

Methodology. We tested ChatGPT on 2026-08-04 using a fabricated candidate profile with no numeric data, so any figure in the output was provably generated. Six runs: three unconstrained, three with the anti-fabrication instruction. Each ran in an isolated session with no memory of the others and no custom instructions. Limitations: six runs is directional rather than precise, no rate here should be read as a population statistic, one model was tested, and no job posting was supplied — so tailoring behavior is untested.

Sources

• Jobscan — Can ATS Detect AI Resumes in 2026?
• ResumeBuilder.com — 3 in 4 Job Seekers Who Used ChatGPT to Write Their Resume Got an Interview
• Harvard College Resumes & Cover Letter Guide — Mignone Center for Career Success