ChatGPT is a good resume editor and an unreliable resume author. Give it your real resume and a job posting, and it will pull out the requirements, reword your bullets in the posting's vocabulary and cut a long draft to one page faster than you can. Ask it to "write" or "improve" a ChatGPT resume with no rules, and it will add metrics you never measured, tools you never used and titles you never held, all in the same confident tone. The fix is to break tailoring into small steps with strict rules, then check the output against your original before anything goes out. This post gives you a prompt for each step and the fact check to run at the end.
What ChatGPT is good at on a resume
The useful work on a resume is mostly language work, and language is what these models do well. Four tasks in particular:
- Reading a posting for what it asks. Postings bury the requirements among mission statements and benefits. Extracting a clean list is fast and accurate, because every answer is sitting in the text you pasted.
- Matching your words to theirs. Your bullet says "late shipments by region"; the posting says "on-time delivery". Spotting that the two describe the same thing, and suggesting the posting's term, is a strength.
- Cutting. Shortening a long bullet without losing the point is tedious for a person and easy for a model.
- Consistency. Tense, verb choice, parallel structure, the small things that make a page read as edited. For spelling and grammar alone, see what Grammarly fixes and misses on a resume.
Notice what these have in common: the facts are already in the conversation. The model is rearranging and rewording, not supplying. That is the zone to keep it in.
Where it goes wrong
OpenAI is direct about the risk. Its help center article Does ChatGPT tell the truth? says ChatGPT "can produce incorrect or misleading outputs", lists "fabricated quotes, studies, citations or references to non-existent sources" among the errors to watch for, and recommends treating what it writes as "a first draft, not a final source". OpenAI's research post Why language models hallucinate defines hallucinations as "plausible but false statements generated by language models" and explains that training rewards guessing over admitting uncertainty: "Strategically guessing when uncertain improves accuracy but increases errors and hallucinations."
On a resume, a plausible guess looks like this:
| What you gave it | What it may return | Why it is a problem |
|---|---|---|
| "Cleaned vendor data files with Python" | "Automated vendor data pipelines in Python, reducing processing time by 40%" | The 40% does not exist |
| "Data Analyst" | "Senior Data Analyst" or "Analytics Lead" | A title is a fact a background check can verify |
| "2021 to 2023" and "2023 to present" | One role, "2021 to present" | Merged dates hide a job change |
| A posting that asks for Snowflake | "Built models in Snowflake" | A tool you may never have touched |
| "Worked toward PMP" | "PMP certified" | A credential you do not hold |
Some popular advice invites this. Coursera's ChatGPT resume guide, updated 11 October 2026, includes a tailoring prompt that asks ChatGPT to "suggest four resume achievements I can incorporate into my most recent experience". The same article warns that "ChatGPT may suggest things about you that aren't true." Both are right, and the second is the reason not to use the first. A model asked for achievements will produce achievements, whether or not they happened.
The temptation is real for job seekers too. In Huntr's survey of 1,049 job seekers for its 2025 job search trends report, 458 people, about 44%, said they would use AI to exaggerate their qualifications. That is a self-reported answer from one vendor's survey, not a measure of what anyone did. But it shows how thin the line is between "make this sound better" and "make this up".
Before you start: set up the chat
A little setup makes the prompts below work better.
- One posting per chat. Start a fresh conversation for each job, so details from one posting do not leak into another. A temporary chat also keeps it out of ChatGPT's memory. OpenAI's data controls page says temporary chats do not create or update memories and are not used to improve its models, though they may be retained for up to 30 days for safety purposes. The same page explains how to turn off "Improve the model for everyone" in Settings, then Data controls.
- Strip your contact details. The model does not need your phone number, address or email to edit bullets.
- Paste your master resume. The long version with every role and bullet, not your last tailored copy. The model can only select from what it sees.
Prompt: pull requirements from a posting
This first prompt does the analysis and nothing else. The key instruction is the last column: the model must quote your evidence or admit there is none.
Below are a job posting and my resume.
Step 1. Ignore the company description, benefits, pay and equal opportunity
text. From what is left, list every requirement: named skills and tools, years
of experience, education, responsibilities, and domain terms.
Step 2. Return a table with these columns:
- Requirement, in the posting's exact words
- Required or preferred (as the posting says; write "unclear" if it does not say)
- Evidence: quote the exact bullet from my resume that proves it, word for word.
If nothing in my resume proves it, write NONE. Do not paraphrase my resume
and do not infer experience I have not written down.
Do not rewrite anything yet.
JOB POSTING:
[paste]
MY RESUME:
[paste]
Read the NONE rows carefully. Each one is a decision for you, not the model. Either you have the experience and left it off your master resume (add it yourself, in your own words), or you do not, and it stays off. This table is also the input for the next two prompts, so keep it in the chat.
Prompt: rewrite bullets in the posting's words
Now the rewrite, with the rules that keep it honest written into the prompt. This is the ChatGPT prompt to tailor a resume to a job description that most guides skip the rules on.
Using the requirements table above, rewrite only the resume bullets that appear
in the Evidence column.
Rules:
1. Where my bullet describes the same thing as the posting, use the posting's
term for it.
2. Keep every number, percentage, dollar amount, employer name, job title,
date, school, degree and certification exactly as I wrote it. Do not add any
new number.
3. Do not add a tool, skill or technology that is not already in that bullet.
4. Do not change the scope: monthly stays monthly, "helped" stays "helped",
a team of two stays a team of two.
5. Lead each bullet with the part the posting cares about most.
Return a two-column table: my original bullet, your rewrite. After the table,
list any requirement marked NONE that you were tempted to work in, so I can
decide about it myself.
The side-by-side table is the point. You will see every change against its original, which makes drift visible at a glance. The final list does something useful with the model's urge to fill gaps: it puts them in front of you as questions instead of slipping them onto the page as claims.
If a rewrite fails a rule, say which one: "Bullet 3 added '40%'. Rule 2. Redo it without the number." Models follow a cited rule better than a general complaint. If you use Claude instead, the Claude resume prompts post runs the same loop and makes Claude quote a source for every edit.
For a full worked example of this step done by hand, with a before and after table, see our guide to tailoring a resume.
Prompt: tighten to one page
Once the bullets are right, length is the last job.
Assemble a one-page resume from my master resume, using the rewritten bullets
above where they exist.
- Keep at most [4] bullets for my current role and [3] for each earlier role,
choosing the bullets that prove the most required items in the table.
- Drop bullets that prove nothing in the table.
- Each bullet should be under [25] words.
- Order the skills line so the posting's named tools come first, using the
posting's spelling, and include only skills that already appear in my resume.
- Do not merge roles, drop dates, change job titles or remove an employer.
Every role keeps its own line with employer, title and date range.
Then tell me which bullets you cut and why.
Set the numbers in brackets to suit your history. Someone with one role and fifteen years in it needs a different split from someone with five roles. If you want a summary line at the top, ask for one in a separate message, with the same rule: it may only use facts already on the page. Our collection of resume summary examples shows what a good one looks like.
A fact check to run before you send
Do not ask the model whether it invented anything. It is checking its own work with the same habits that produced it. Check the facts yourself, against your master resume, using a method that does not depend on trusting the output.
1. Extract, then verify by search. Ask for a list, which is a low-risk task, then check each item yourself:
From the final resume, list every number, percentage, dollar amount, date,
job title, employer, school, degree, certification and named tool, one per line,
exactly as written. Do not comment on them.
Open your master resume and search for each item with Ctrl+F (Cmd+F on a Mac). Anything you cannot find word for word gets investigated. Most will be harmless rewording; the ones that matter are new numbers, new titles and new tools.
2. Check the skeleton by eye. Put the two versions side by side and confirm for each role: same employer, same title, same start and end dates. This takes a few moments and catches merged or stretched roles, which the list above can miss.
3. Run the interview test. Read every skill and every bullet and ask whether you could talk about it for two minutes with someone who knows the subject. Anything that fails comes off, however well it matches the posting.
4. Check how it parses. Paste the final text into your resume template, export it, then open the file, select all, copy and paste into a plain text editor. Your name, roles, dates and keywords should all come through as real text in a sensible order.
| Check | What it catches |
|---|---|
| Extract and search | New numbers, tools, titles, credentials |
| Skeleton by eye | Merged roles, changed dates, upgraded titles |
| Interview test | Claims you cannot support |
| Parse test | Formatting that hides text from the system reading it |
When to automate instead
The workflow above works, and it is free if you already use ChatGPT. Its cost is your time, and the time scales with every application. Three prompts, the corrections, the fact check and the formatting add up per posting, and Huntr's 2025 report put the median job seeker at about 16 applications a week. At that pace the careful version of this process becomes a second job, and the fact check is the step people drop first.
That is the case for tailoring tools that enforce the rules in code rather than in a prompt. ApplyEze is one; our comparison of the best AI for resume tailoring covers others and where each draws its lines. ApplyEze rewrites every resume for the posting in front of it and writes a cover letter when the form asks. It mirrors the posting's own vocabulary, the words a recruiter is likely to search for, and it brings the posting's named skills and tools onto your skills line, even when your own record does not list them. There is no step to remove one before a live run, so use a dry run to see what it writes, and be ready to discuss anything on the page. A certification the posting asks for is not added unless you hold it. Your employers, job titles, dates, schools, degrees and certifications come from your own record, and code rejects a rewrite that adds a figure, employer, school or credential your record does not hold. A prompt can ask a model not to invent a metric; a code check refuses the rewrite when it does.
If you edit by hand, list only skills you can discuss in an interview, and read the tailored resume before it goes out. Whichever route you take, it is your name on it.