Which Claude model is best for resume writing?

Which Claude model should you pick for resume writing?

Claude Opus 5.5 is the best Claude model for resume writing, and the gap is widest on the hardest part of the job: turning a messy ten-year history into fifteen bullets that still sound like you wrote them. Opus holds more of your material in working context at once, so when you paste a three-page CV, the job ad and last year's performance review together, it reasons across all three rather than fixating on whatever you pasted last. Resume writing is a judgement task before it is a writing task. Deciding which of your four projects earns a line and which gets cut is precisely where the lighter models go vague and start producing the same five bullets for everybody.

Sonnet 5.5 is the workhorse, and for most job seekers it is the sensible default. If you already have a decent CV and you are tailoring it to a specific advert, Sonnet produces rewrites that are close enough to Opus that you would struggle to pick them apart in a blind test, and it turns them round fast enough to run eight variations in an evening. Haiku 5.5 sits further down again: use it for mechanical work, like converting a wall of paragraph text into bullet form, standardising your date formats, or trimming a bullet from 34 words to 20. Ask Haiku to make strategic calls about positioning and you will get filler.

Fable 5.1 is the odd one out, and it earns its place on a narrow job: voice. If every draft you produce reads like a committee wrote it, Fable is the model to hand your rough, honest, badly written description of what you actually did, and ask it to keep your phrasing while tightening the grammar. What it should not do is decide your structure or your keyword coverage, because a CV that reads beautifully and buries the job title the recruiter is scanning for still loses. My practical rule: Fable for the summary paragraph and nothing else, Opus for the architecture, Sonnet for the twenty tailoring passes that follow.

Is Claude good at resume writing, or does it just sound confident?

Claude is genuinely good at two resume jobs and genuinely bad at a third. It is strong at compression, taking a rambling paragraph about what you did and reshaping it into a tight, active bullet, and it is strong at translation, converting internal jargon into language a recruiter outside your company understands. Where it fails is evidence. Claude cannot know that your migration cut deploy time by 40 per cent, so if you do not give it the number, it will either leave a hole or quietly write something plausible. That second outcome is the dangerous one, because a fabricated figure on your CV is your problem at interview, not the model's.

The other weakness is sameness. Ask any Claude model for a professional summary with no constraints and you get the same architecture every time: a seniority label, three adjectives, a domain, a promise about driving impact. Recruiters read hundreds of these a week and they have learned to skim straight past them. The fix is to feed the model raw material it could not have guessed, specific tools, specific numbers, specific friction you resolved, and to tell it explicitly not to add anything you did not supply. If you want the full picture of where each major assistant helps and where it quietly hurts, the head-to-head on ChatGPT and Claude for CVs walks through both.

There is also a rules question worth settling before you start, because job seekers keep asking it. Using an AI assistant to rewrite your own experience is not deception; inventing credentials, employers or results is, and that line does not move depending on which model you used. If you want the detail on where the boundaries actually sit, see the breakdown of what Claude's usage policy says about AI-written CVs. Keep your drafting honest and the model becomes what it should be: a fast editor with no ego, happy to produce the eleventh version of a bullet at midnight without sighing at you.

What Claude resume prompts actually produce usable bullets?

The prompt that works puts the evidence in and the invention out. Start every resume prompt with a hard constraint along these lines: rewrite only from the facts I give you, do not add numbers, tools, employers or outcomes I have not supplied, and flag any bullet where you think a metric is missing. That single instruction removes most of the fabrication risk and turns the gaps into a to-do list instead of a liability. Then give the model its role and its reader. Telling it that it is editing for a hiring manager who will spend eleven seconds on the page produces much sharper output than asking for a professional rewrite, because it forces decisions about what goes first.

For tailoring, the structure that pays off is three-part: paste the full job advert, paste your current CV, then ask for a gap list before any rewriting happens. Make the model tell you which requirements in the advert your CV does not currently evidence, ranked by how central they are to the role. Read that list, decide honestly which gaps you can close with real experience you simply forgot to mention, and only then ask for the rewrite. People who skip the gap step end up with a CV that mirrors the advert's vocabulary without proving anything, which is the exact failure mode experienced recruiters spot fastest.

For individual bullets, give the model the shape you want rather than hoping it guesses. Ask for the action, the method and the measurable result in under 22 words, active voice, no adverbs, past tense for finished work and present tense for your current role. Then ask for three variants at different levels of specificity so you can pick rather than accept. If you would rather not run this loop by hand for every application, a chat-based CV builder does the same thing with the structure already enforced, and exports into templates that hold their formatting in PDF and Word.

When is the cheaper Claude model good enough for your CV?

Sonnet 5.5 is good enough the moment your CV's structure is already settled. Once you have decided what goes on page one, which roles get four bullets and which get one, and what your summary claims, everything left is editing work, and the cheaper model does editing work well. The split I would hold to is simple: pay for reasoning, not for typing. One Opus session to build the base CV, then Sonnet for every application after that. Over a two-month search with thirty applications, that is one expensive run and twenty-nine cheap ones, and the quality difference on the tailoring passes is small enough that it rarely decides an outcome.

Haiku 5.5 earns its keep on the unglamorous jobs nobody wants to do by hand. Normalising inconsistent date formats across eight roles, converting American spellings to British ones or the reverse, cutting a 640-word experience section to 480 without losing a fact, checking that every bullet starts with a verb. These are rule-following tasks with a verifiable answer, which is exactly what a fast, cheap model is for. The mistake is reaching for Haiku when you are stuck on what to say rather than how to say it. If you find yourself re-prompting three times because the output is bland, you have picked the wrong tier, not the wrong wording.

There is one case where I would spend the money without hesitating, even for a single application: a career change or a return after a break. Those CVs need the model to understand two different worlds at once and build a bridge between them, reframing classroom work as training delivery or clinical coordination as project management. That is reasoning under constraint, and lighter models tend to either keep your old vocabulary or lose the substance entirely. The same goes for senior roles where the CV has to show scope and decisions rather than tasks. Everywhere else, the cheap model and a good prompt will get you most of the way.

How do you check Claude's resume output survives an ATS?

Check the file, not the text. Whatever Claude produces in the chat window is plain text that reads perfectly; what the employer's system receives is a PDF or Word document whose layout decides whether your job titles and dates come out in the right order. The two things that break parsing have nothing to do with the writing: multi-column layouts, which many parsers read straight across and scramble into nonsense, and text locked inside tables, headers, footers, text boxes or images, which often does not get extracted at all. A beautifully worded CV in a two-column template can arrive at the recruiter's screen with your employer names interleaved with your skills list.

So the last step of any AI-assisted CV session is an extraction test. Upload the finished file to a checker and look at what comes back out: are your dates intact, are the job titles attached to the right employers, did the contact details survive, is anything missing entirely? HRLens runs this free, and gives you a score out of 100 with five category scores covering experience, tech stack, impact and ownership, clarity and structure, and ATS compatibility, plus a visual layout analysis showing how the page reads. You can run your CV through the free analysis before you send a single application.

Do the same check on every exported variant, not just the first. If you tailor with Sonnet across fifteen applications and export each one, a template change or an inserted icon can break a file you assumed was safe. Paste the job description into the analysis too; it will list the skills from that advert your CV does not currently evidence, which is the same gap list a good prompt produces, except it comes from the parsed file rather than the chat transcript. That is the honest division of labour: Claude writes the words, and a parse test tells you whether anybody is ever going to read them.

Frequently asked questions