How to Write a Resume with AI in 2026 (Step by Step)

Why AI resumes fail more often than they succeed

The problem is not the model. GPT-4-class systems, Claude Opus, and Gemini Advanced can all write competent resume bullets when prompted correctly. The problem is the workflow most candidates use: open the AI, paste the role title, take whatever comes back, paste it into a template, and ship. The output reads like every other resume that came out of the same prompt that day. Recruiters who screen 200 resumes a week pick up on the pattern by Tuesday.

The fix is not to abandon AI. It is to use AI for what it is good at, which is converting structured notes into clean, scannable copy, and to keep the human work where it belongs, which is choosing the right experiences, the right numbers, and the right level of specificity for the target role. The five steps below are the workflow we use internally at Curriq, the AI resume builder for iPhone, and the workflow we recommend to anyone we coach.

Step 1: Brain-dump every role into structured notes

Before you open the AI, write a flat list of every role you want to consider for inclusion. For each role, capture six fields: company, title, dates, scope (team size, budget, region, or product surface you owned), stack (tools, languages, frameworks, methods), and three outcomes. Outcomes are the part most candidates skip. They are not responsibilities. They are things that changed because you were there.

A useful test for an outcome: if a coworker reading the bullet would say "wait, that was you?", it is a real outcome. If they would say "that is just what your team does", it is a responsibility, and responsibilities do not belong on a resume in 2026. Numbers are a force multiplier here. If you saved 8 hours a week, ran a 40-person team, deployed across 14 markets, or shipped 6 features per quarter, write the number now. You will use it as the anchor for the bullet, not as a decoration on top of it.

Spend 20 minutes on this step. The temptation is to cut it short and make the AI do the recall, but the model has no access to what you actually did. Whatever you do not capture here, the AI will fill in with a generic placeholder, and that placeholder is the first thing recruiters spot.

Step 2: Feed the AI the target job description first

Open a fresh chat. Paste the full job description for the role you are targeting before you paste anything else. Then add one sentence of instruction: You are helping me draft a resume for this role. I will give you my background next. Confirm the 6 most important keywords from the job description before we start, in priority order.

The reason for this step is that language models anchor on whatever appears first in the context window. If you paste your background first, the model writes generic resume bullets and tries to bend them to the job. If you paste the job description first, the model treats it as the goal and writes bullets that map your background to the role's vocabulary. The difference is visible in the first draft.

The keyword confirmation step is a small but useful guardrail. It surfaces the model's interpretation of the role before it writes a single bullet, so you can correct it. If the model says the top three keywords are "stakeholder management", "agile", and "cross-functional collaboration", you have a vague job description and you should pick a different posting to anchor against. If it says "Snowflake", "dbt", and "PII compliance", you have a useful target and you can move on.

Step 3: Generate bullets with impact-action-method

Now paste your structured notes. For each role, ask the model to produce three to five bullets in the structure outcome + verb + method. The outcome comes first because it is the only part the recruiter cares about. The verb is the action you took. The method is the how, the proof that you did the thing.

A weak bullet: Responsible for managing a team of analysts and improving reporting workflows. An impact-action-method bullet: Cut weekly reporting time from 14 to 3 hours by leading a 4-analyst team to migrate 22 dashboards from Tableau to Looker with a shared dbt model. Same role, same person, completely different signal.

If a bullet has no number, that is fine, but it has to have a verifiable outcome. Wrote the API contract that all 3 product teams use today is verifiable. Drove cross-functional alignment is not. Tell the model explicitly: do not invent numbers. If a metric is not in my notes, write the bullet without one. Resumes get verified during reference checks, and a fabricated metric is the easiest way to get an offer pulled.

Step 4: Cut the AI tells

The single biggest reason AI-written resumes get spotted is a small set of words and phrases that the major models over-use. Search-and-destroy these from the draft before you do anything else:

• Leveraged, spearheaded, orchestrated, championed, fostered. Replace with led, built, shipped, cut, grew.
• Passionate about, dedicated to, highly motivated. These are unverifiable. Cut.
• Cross-functional collaboration, stakeholder alignment. These are abstractions. Replace with the actual names: worked with sales, ops, and finance beats collaborated cross-functionally.
• A track record of, proven ability to. Resumes are evidence, not claims. Cut.
• Any opener that starts with Results-driven professional or Dynamic leader. The summary section should be three lines: who you are, what you do, what you want next. Plain English.

Then look for sentence-level tells. AI loves the construction X to drive Y (used dbt to drive reporting velocity). Recruiters call this the consultant voice. Cut the to drive. Used dbt to cut reporting time from 14 to 3 hours is the same bullet, but specific. AI also loves the suffix resulting in. Cut it. Lead with the number: Cut reporting time from 14 to 3 hours by migrating 22 dashboards to Looker.

Step 5: Run an ATS format check before export

The ATS (applicant tracking system) at most companies in 2026 is one of about 30 mainstream products. They share most of their parsing rules. The format choices that break parsing are the same across all of them:

• One column. Two-column resumes look elegant on the screen and parse as a wall of overlapping text in the ATS.
• No text boxes, no tables. If your name and contact info are in a table, half of scanners will lose them.
• No icons in section headers. The little envelope before your email, the little phone before your number, the little house before your address. They render as glyphs the ATS cannot read, and on bad scans they corrupt the line.
• Standard section names. "Experience", "Education", "Skills". Not "My Journey", not "What I Bring".
• Embedded fonts. If you used Inter, Outfit, or any non-default font, embed it on PDF export, or fall back to Arial / Calibri / Helvetica.
• US Letter or A4. Pick one based on the country of the role. US, Canada → US Letter. UK, EU → A4.

None of this is new. What is new in 2026 is that LLM-based parsers are starting to ship inside ATS platforms, and they are far more forgiving on the format side. The catch is that you do not know which version a given employer is running. Default to the strict format and you cover both.

The 9 mistakes that get AI-written resumes spotted

In rough order of how often we see them:

1. Bullets with the words leveraged or spearheaded in the first three lines.
2. A summary that says results-driven [role] with a passion for.
3. Identical sentence structures across every bullet (every one starts with a verb, every one ends with resulting in).
4. Numbers that are suspiciously round (50%, 100%, 200%) on every bullet, when most real outcomes are 37% or 14x.
5. Skill lists that include 28 items, half of which are buzzwords like strategic thinking.
6. The same outcomes repeated under three different roles ("improved team productivity").
7. A cover letter that opens with I am writing to express my interest in.
8. Education listed before experience for someone who has been working 5+ years.
9. A "Soft Skills" section. Nobody puts soft skills on a resume in 2026.

What to do when the AI loops

Models sometimes get stuck. They produce three drafts in a row that say almost the same thing in slightly different words. When that happens, change the input, not the prompt. Add a constraint that forces a different angle: rewrite this bullet from the perspective of the customer who benefited, or rewrite this bullet starting with the dollar amount and let the rest follow. Constraint changes break the loop faster than instruction changes.

If you are using Curriq specifically, the in-app rewrite has four built-in modes (Professional, Confident, Creative, Conservative). Switching modes is the fastest way to break a loop. The same bullet in Conservative mode is a calmer version, in Confident mode owns the impact more directly, in Creative mode tries a fresh framing. None of them invent metrics; the schema prevents it.

The end-to-end check before export

Before you export the PDF, walk through this list. It takes 4 minutes:

1. Read the resume top to bottom in 30 seconds. If anyone could have written it, rewrite the first paragraph.
2. Verify every number against your notes. Unverifiable numbers come out.
3. Search for the words leveraged, spearheaded, passionate about. Replace any hits.
4. Check that every bullet has either a number or a verifiable outcome.
5. Confirm one column, no icons in headers, embedded fonts, correct page size.
6. Export to PDF. Open it in a different app to confirm it renders correctly. Some apps emit broken kerning or missing fonts at export time.
7. Send a test copy to your own email. Open it on your phone. If anything is mis-laid out on mobile preview, fix it before sending.

What changes in 2026 vs earlier years

Three real shifts. First, recruiter tooling is starting to flag AI-written resumes by phrase pattern. The defense is exactly what step 4 above describes: cut the giveaways. Second, ATS platforms are slowly adopting LLM-based parsing, which forgives more layouts. The right move is to keep the strict format anyway, because the bottom-half of employers will run the older parsers for years. Third, candidates are routinely sending the same AI-written resume to 100+ jobs, which makes recruiters less patient with generic-sounding bullets. Tailoring matters more than it did in 2023.

For the tailoring workflow specifically, see our deeper guide on how to tailor a resume for a specific job posting. For the ATS-format details, see ATS-friendly resume templates that pass scanners.