Can you build a CV just by chatting with AI?

Can you really build a CV just by chatting with AI?

Yes — you can build a complete CV through a back-and-forth conversation, and the result holds up, provided you supply the raw material and strike out the invented numbers. The format just picked up a loud endorsement from an unlikely direction. On 1 October 2026, Shopify launched Canvas, which lets merchants build and redesign their online stores by chatting with its AI agent, Sidekick, watching the real storefront code re-render as they talk rather than a static mockup. Sidekick asks questions, proposes a direction, takes feedback and remembers earlier choices so you aren't repeating yourself. Swap "storefront" for "CV" and you have the same interaction model job seekers have been improvising inside chatbots for two years.

The reason it works is the blank-page problem. A traditional template hands you empty fields and no idea what good looks like inside them, so most people type their job description into the boxes and call it a CV. A conversation does the opposite: it interviews you, and interviews are how facts surface. That's exactly how the HRLens CV builder is set up — paste an old CV or simply describe what you've been doing, refine it by chat, then pick from six templates with PDF and Word export. The questions carry the structure, so you're never staring at a cursor wondering what belongs under "profile".

Here's the honest limit, and it's worth saying plainly: chat is a brilliant interviewer and a mediocre fact-checker. Canvas itself ships with edges — desktop-only at launch, no support yet for third-party themes or translations, and it doesn't replace the existing editor. Conversational CV building is the same. The chat gets you to a strong draft faster than any form-filling tool, but it will also happily hand you a confident sentence built on nothing, and it has no idea whether the file you export survives the parser on the other end. Those two gaps are where the rest of this article lives, because they're where chat-built CVs actually fail.

What should you paste in before you start chatting?

Paste three things before you answer a single question: your most recent CV as plain text, the full text of the job ad you're targeting, and a messy list of numbers from the past two or three years. That trio decides the quality of everything downstream. A builder working from nothing can only ask generic questions, so you get generic bullets back. Hand it your old CV and it already knows your titles, dates and employers, which lets the conversation skip the admin and go straight to what you achieved. Dates matter more than people expect — gaps, overlaps and contract stints are precisely where parsers stumble, and they're far easier to sort out in the first message than the twentieth.

Paste the job ad as text, not as a link. LinkedIn and Indeed block automated fetching, so a URL often arrives as an empty string and the assistant quietly invents a job description to work from. With the real text in front of it, a builder can tell you which required skills your CV never mentions and mirror the employer's own phrasing — "stakeholder management" instead of your "worked with teams". That's what the job-targeting step does at HRLens: you paste the description, the analysis lists what's missing, and the rewrites come back echoing the ad's keywords across six ATS-friendly templates. One ad per session, though. Tailoring to three roles at once produces a CV that fits none of them.

The numbers list is the part almost everyone skips, and it's the part that makes a chat-built CV defensible. You don't need polished sentences, just scraps: team size, the budget you signed off, monthly ticket volume, pipeline value, average deal size, uptime, cohort numbers, how many sites you covered, how long a release cycle took before and after you touched it. Old performance reviews, dashboards, invoices, commit history and your own calendar beat memory every time. Where you genuinely can't retrieve a figure, write the scale in words — "a three-person team inside a 40-person department" — and tell the assistant not to guess. Supplying real data is the one job no model can do on your behalf.

Which follow-up prompts turn vague answers into quantified bullets?

Four follow-ups do most of the heavy lifting: compared to what, how was it measured, what did you do versus the team, and what happened afterwards. Put those to every bullet and duty statements turn into evidence. Most first drafts from a chat builder read like a job description because that's what you fed it — "responsible for managing the reporting process" describes a seat, not a contribution. The baseline question changes the most. "Improved reporting" means nothing standing alone; "the monthly close used to take nine days" gives the reader somewhere to stand. Recruiters read the distance between before and after, not the adjective you parked in front of it.

The measurement question is your fabrication filter. If you can't say where a number came from — which dashboard, whose report, what period — it has no business on the page, and a hiring manager will establish that in the first ten minutes of a call. Follow-up questioning is still how inflated figures get caught: what was the baseline, how was it tracked, what else changed at the same time. Answer those in the chat and you've rehearsed the interview as a side effect. The ownership question handles the other common distortion. "Led a migration" is doing a lot of quiet work; "owned the data model and cutover plan on a twelve-person migration" is smaller and far more believable.

The fourth question — what happened six months later — separates a launch from a result. Plenty of projects spike then sag, and the ones that held deserve the extra line. Once the substance is right, hand the assistant hard constraints instead of vague encouragement: one line per bullet, roughly 18 to 25 words, a different opening verb each time, past tense for finished work and present for the current role, no adjectives about yourself. Ask for before-and-after pairs so you can see what it changed and reject half of it. Then read the bullets aloud. Anything you'd squirm defending across a table, cut it right there, while the conversation is still cheap.

Where do chat builders invent numbers you have to delete?

Chat builders invent in three predictable places: percentage improvements, team sizes and budgets, and tools you never touched. Watch for the giveaway round number — "increased efficiency by 30%", "reduced costs by 25%", "managed a team of 10" — which shows up because the model is pattern-matching strong CVs rather than reading your life. The third category is sneakier. Describe a data role and a generous assistant slides in SQL, a scheduler and a dashboard tool as a reasonable inference. Mention customer support and a helpdesk platform appears. None of it is malicious; the model fills gaps the way autocomplete does. Your job is to treat every specific it produces as a question, never as a statement of fact.

This matters more than it did two years ago, because screeners now assume AI was involved somewhere. Recruiter surveys through 2025 and 2026 consistently report that a large majority — roughly seven in ten — have encountered applications with fabricated histories, invented references or plainly machine-written copy, and many hiring managers say AI polish has made real skills harder to verify. The response hasn't been to ban AI. It's been to probe harder. What actually gets a CV binned isn't that a model helped write it, but that the writing is generic, interchangeable and unverifiable. Bullets all cut to the shape "Led X, resulting in Y% improvement in Z" are the pattern people have learned to distrust, because real work is lumpier than that.

So apply a blunt rule: any number you didn't type yourself gets deleted. Not softened, not hedged with "approximately" — deleted. Where a bullet feels naked without a figure, substitute scope instead: the headcount you coordinated, the systems you were on call for, the territory you covered, how many releases you shipped. Scope is verifiable and it doesn't buckle under questioning. It's also worth checking how the finished draft lands with a sceptical reader, because the giveaways tend to be phrasing and rhythm as much as suspicious statistics — a quick check for whether your CV reads as AI-written will flag the lines that sound generated before a recruiter has the chance to.

When should you stop chatting and export to PDF or Word?

Stop chatting the moment your next message would be cosmetic rather than factual. Three signals say you're finished: every bullet carries a number or a concrete scope you could defend out loud, the skills named in the job ad appear in your own words, and the assistant has started offering synonyms instead of asking questions. Conversational tools are superb at drawing information out of you and poor at recognising when enough is enough — left running, they'll keep proposing punchier verbs until the whole document reads like an awards submission. Shopify's new builder has the same shape and the same boundaries, launching desktop-only and without third-party themes or translations, which is a useful reminder that these tools are drafting partners, not finishers.

Then export, and let the application decide the format. Modern applicant tracking systems read both PDF and Word without complaint, so the extension matters far less than the layout underneath it. Word remains the safer default for online portals, especially stricter legacy engines like Taleo and iCIMS, where extraction accuracy slips on anything structurally complex. PDF is the better pick for emailing a recruiter directly, since it arrives looking exactly as you left it. The genuine killer is neither: an image-based file — a scan, a screenshot, or a design export whose text sits inside graphics — parses as a blank page. Keep both versions from one single-column source, and follow whatever the portal explicitly asks for.

Before you send anything, read the exported file rather than the chat window; they are not the same document. Open the PDF, select all, copy, and paste it into a plain text editor — what appears is roughly what a parser sees. Scrambled dates, vanished headings or a merged column are all fixable while it's still a draft. A free CV analysis takes it further, returning a score out of 100, five category scores including ATS compatibility, your strengths and a visual layout check on the actual file. Applying in Israel? The builder writes Hebrew CVs in Hebrew and exports them right-to-left, which is exactly where extraction usually breaks.

Frequently asked questions