Make a Resume With AI: The File Recruiters Actually Read

Key takeaways

Try our awesome AI for free

Ask me anything...

No single file format is accepted by all ten hiring systems we checked. PDF works with nine of them and DOCX works with nine too, but not the same nine: SAP SuccessFactors never mentions PDF on its parsing page, and Welcome to the Jungle accepts nothing but PDF.

Published size caps range from 1 MB to 100 MB depending on the vendor. Oracle Taleo stops at 1 MB by default, LinkedIn recommends staying under 2 MB, Welcome to the Jungle and Workable cap at 5 MB. Greenhouse accepts up to 100 MB but writes that above 2.5 MB it no longer reads the content.

The resume produced by a free AI resume builder comes out in two columns, exactly the layout Greenhouse lists among the causes of a failed parse. Tested end to end, file downloaded and opened.

Measurements taken on August 10, 2026 from the official documentation of ten vendors, with screenshots.

Making a resume with AI takes ten minutes. Getting the machine that receives it to actually read it is another matter, and it is the part nobody covers with numbers. The pages ranking for this query all repeat the same common sense advice about columns and tables, without ever naming a single system or quoting a single line of documentation.

We did the opposite. We opened the official help pages of ten systems that receive resumes, recorded what they really accept, then generated a resume with a free AI builder and opened the file it produced. Here is what came out, along with the writing method that goes with it.

What AI writes well, and what it cannot write for you

A language model is good at three things on a resume: turning a vague duty into an active sentence, finding the right verb, and matching vocabulary to a job posting. It is bad at exactly one thing, and that one is decisive: it does not know your career. Anything it adds that you did not give it is an invention, and an invention on a resume gets paid for in the interview.

The method in five steps

The prompt to copy

This template has five blocks: role, material, posting, constraints, output format. Replace what sits inside brackets and leave the rest alone.

You are a recruiter in the [industry] industry. You are reviewing a resume
before it goes out.

MY RAW MATERIAL (everything I give you is true, add nothing) :
- Role : [exact title] at [company], [city], from [month year] to [month year]
- What I did : [3 to 5 accomplishments, with a number whenever I have one]
- Education : [degree, school, year]
- Tools : [software, languages, certifications]

THE JOB POSTING :
[paste the full text of the ad here]

WHAT I WANT :
1. A summary of 3 lines maximum, written without "I", that picks up the
   exact job title from the posting.
2. For each role, 3 bullets that start with an action verb in the past
   tense, one idea per bullet, and one number per bullet whenever my
   material contains one.
3. The list of words from the posting that appear nowhere in my material,
   as a two column table : the word, and why it is missing.

CONSTRAINTS :
- Invent no number, no date, no name, no result. If something is missing,
  write [TO BE FILLED] instead.
- No superlatives, no "passionate about", no "dynamic".
- Return plain text only. No layout, no columns, no tables outside item 3.

The constraint that matters is the second to last one. Without it, a model fills the blanks. With it, it hands you the list of what you are missing, which is far more useful than a smooth piece of text.

Before and after

The numbers in the right hand column do not come from the model: they come from your raw material. The model only put them in the right place, in the right shape. That is exactly the division of labor to aim for.

What hiring systems really accept

A resume submitted online almost always goes through an applicant tracking system before it reaches a human. That system starts by extracting the text from your file to fill in a candidate record. If it fails, your application still exists, but it arrives empty, and it is up to the recruiter to retype it by hand.

On August 10, 2026 we recorded, from the public documentation of ten vendors, the accepted formats and the published size cap. The table below is the database behind the tool that follows.

Checked on August 10, 2026

The same page notes that SAP does not do the job itself: “SAP SuccessFactors uses the third-party software Textkernel to parse resume data.” Several vendors outsource resume parsing this way, which explains why their format lists diverge with no obvious logic.

These findings reflect what vendors publish, not what each employer has configured. An Oracle Taleo customer can raise the 1 MB cap and an SAP customer can enable other formats. The documentation gives the default value: that is what you run into when nobody has changed anything, so in most cases.

Test your file

The tool below applies the table row by row. You can hand it your resume: your own browser reads it locally, nothing is sent anywhere, and only the extension and the size are measured. You can also type everything in by hand.

What Greenhouse says it rejects, word for word

The Greenhouse help page devoted to failed parses is the most explicit of the set. It lists the layouts that break extraction. Here is that list, quoted exactly.

• “Resumes that are uploaded as an image, rather than a document (such as a .docx or .pdf)”;
• “Complex resumes with tables, headers, and footers”;
• “Resumes with the name and contact information in the header, footer, or text box”;
• “Resumes that have a columned layout”;
• “Resumes without clear sections and differing formats throughout each section”;
• “Resumes that include graphics, photos, or word art”;
• “Resumes with incomplete job titles”, with “Sr. Account Exec” instead of “Senior Account Executive” given as the example.

Four of those seven points describe exactly what a modern resume template produces: columns, boxes, contact details in a sidebar, icons. That is the core of the problem, and it has nothing to do with the text the AI wrote.

Checked on August 10, 2026

Lever takes the reasoning further and gives the simplest test there is: “A simple test to gauge if a file’s content is parseable is to try highlighting text in the document with your cursor.” In other words, open your PDF and try to highlight your own name. If the cursor slides across without catching anything, your resume is an image, and none of the systems in the table will read it, except SAP SuccessFactors, which does document parsing image files.

What an AI builder really produces

That leaves the question of what the tools promising a resume in three minutes actually put out. We took one, the Adecco Foundation’s CV Maker, for a specific reason: it is free, it asks for no account, and it exists in English as well as in Spanish. In other words, the shortest path between a job seeker and a resume made by AI.

We filled it in with a fictional profile, used the AI assistant for the experience section, downloaded the PDF, then opened the file.

Checked on August 10, 2026

The resulting file has real qualities, and that deserves saying: 87.2 KB, one A4 page, no embedded image, and a single typeface, Roboto Regular. On size it clears every cap, including Oracle Taleo’s 1 MB. On format it clears them too, except SAP SuccessFactors, which does not list PDF. Run through the tool above, the file scores nine systems out of ten, and that is the only blocker.

Then comes the layout.

Checked on August 10, 2026

It is a two column resume, with the name and contact details cut off from the content they are supposed to identify. That is one of the seven points on the Greenhouse list, in one shot.

Eleven lines of text extracted, 199 characters. The start and end dates come out as two lone hyphens, and the company and the city end up glued together on one line. None of this stops a human from reading the document: all of it complicates the job of the program trying to turn it into a candidate record.

The conclusion is not that these tools should be avoided. It is that a resume builder optimizes for looks, not for machine reading, and it never warns you. The text it produces is reusable as is: it is the container that needs rebuilding.

The photo, the icons and the header

Three decorative elements show up in every resume template and each one causes a different problem.

Workday points the same way in its administrator documentation: “Resume parsing results can vary based on resume format and order of words. For best results, use resumes that don’t have images or image-based styles.” The same page notes that two fields are never auto filled, languages and skills, which means a candidate is far better off typing them into the form than trusting the resume to carry them.

Should you say you used AI

The question comes up in every conversation, and it is worth splitting in two. Having used an assistant to draft your resume is not a skill and does not belong on the page. Knowing how to work with AI in your field is a skill, and the guidance on that is public.

Harvard’s Mignone Center for Career Success devotes a page to using generative AI for resumes and cover letters. Its position is blunt: “Generative AI should not be the primary author”. Three of its recommendations are worth copying as they are.

• “Start with your draft.” Begin with your existing resume, or with as many details from your past roles as you can pull together, including the impact you had and the relevant numbers.
• “Use generated text as a suggested edit, not as a final product.” Adjust the language so it is an accurate and authentic depiction of your experience.
• “Be able to speak to every line of your resume if asked in an interview. Read it aloud to catch repetitions or inaccuracies.”

The same page adds a rule that gets forgotten: “Follow organization guidelines. Read all application instructions, as some organizations may have rules about whether AI can be used in preparing application materials.” Reading the instructions before generating anything costs thirty seconds.

Checked on August 10, 2026

The five mistakes that show

For the final pass, two habits that work together: run the text through an AI spelling and grammar checker, then read it out loud. What sounds false out loud will sound false in the interview, where you will be asked to expand on those very sentences.

Frequently asked questions

What to take away

Making a resume with AI works, as long as you know where its job ends. It writes the text, it spots the missing words from the posting, it turns a vague duty into an active sentence. It does not know your career and it does not build a file that the systems receiving it can read.

The move that changes everything fits in one line: drop the text the AI produced into a single column document, with no header, no text boxes, under a megabyte, and check that your name can be highlighted with the cursor. The rest of the work, AI does better and faster than you. To go further on writing prompts, our library of ChatGPT prompt examples also covers the job search.

Primary sources consulted on August 10, 2026: help pages from Greenhouse (accepted formats, failed parses), Lever (resume parsing), SmartRecruiters (application submission), Oracle Taleo (attachment settings, release 22A), SAP SuccessFactors (resume parsing), Workday (resume parsing), Teamtailor, Recruitee, Welcome to the Jungle, LinkedIn and Workable; the Harvard Mignone Center for Career Success page on generative AI for resumes and cover letters. The test resume was generated on the Adecco Foundation CV Maker with a fictional profile, then analyzed offline.