How I rewrote my resume with Claude (and why you should too, right now)

If you’re anything like me, updating your resume sucks. I take a lot of pride in the good work that I perform, but remembering it, capturing it, quantifying it, determining what is worth including, and self-aggrandizing in a reasonable way is a misery. So I went the expected route and asked Claude to help me out. But the approach turned out to be pretty solid so I wanted to share it with you.

I’m an Engineering Manager with 25+ years of experience. I lead a team at Meta which supports the infrastructure for one of the largest dedicated AI/ML research GPU fleets in the industry. By any reasonable measure I should have a resume that opens doors but somehow mine didn’t represent that well at all. Generic summary, undersold bullets, no quantification, the wrong things emphasized. I knew it was bad. I just didn’t want to admit how bad until I got serious about fixing it.

This week Meta is doing a round of layoffs, expected to target around 10% of the company. I don’t know (yet) if I’m going to be impacted, but I do know it’s the responsible thing to get in front of this. Besides that, we’re in an industry-wide reset where being AI-native isn’t optional anymore and we should collectively brace ourselves for more of these. “I used Claude to rewrite my resume” might not be the kind of thing you’d bring up in an interview, but it’s the type of workflow you need to be adopting and considering with each problem you face.

The opening move that mattered

Most people who try to use an LLM for resume work paste their resume in and ask it to “make this better.” You get back a slightly punchier version of the same document. The bullets are tighter, the verbs are stronger, and nothing important has changed. The model is doing local optimization on the input you gave it, and the input is the problem.

The move that worked for me was to start with a diagnostic question instead of a rewrite request.

Prompt that opened the door: “Here’s my resume. I haven’t been getting responses, though I haven’t been applying for much. But I feel like I should get more traction given my history. What should I improve to get past recruiters? If I can get into the phone call I know I’m good, I just need in the door.”

That framing (here’s the symptom, what’s the disease) got me a response that started with “your summary is hiding your best stuff” and walked through six specific structural problems with the document. I had bullets about activity instead of outcomes. My skills section was ATS-weak. Older roles were taking up disproportionate space.

The thing I’d been most proud of, running GPU infrastructure at Reality Labs Research, was buried in the second paragraph of my most recent role. The hook that should have stopped any recruiter scrolling (one of the largest dedicated AI/ML research GPU fleets in the industry) wasn’t anywhere on the page.

Asking for an interview, not a rewrite

After the diagnostic round, we did a first rewrite of the top third of the resume. It was already significantly better. But Claude was making the words sharper, not finding new content. To get past that ceiling, I needed to give the conversation a different shape.

I told Claude I wanted a long-form session structured like an interview. I was going to give it a stream-of-consciousness blob of things that I thought were relevant about my different positions, and it would translate the good parts into bullets and push back on what didn’t belong. Not a single round of revisions. Hours of back-and-forth, role by role, with the model in the driver’s seat asking questions.

Prompt that changed the dynamic: “I’d like for this to be a longer planning session and not one off changes, so you should dig in with questions on anything that is worth exploring, but only if its helpful, we don’t have all day.”

This was the most important framing shift in the whole session. It moved me from “give the AI input, get output back” to “have a working conversation with someone who’s helping me build something.” Claude started asking pointed questions about projects, about scope, about the gaps between what I’d claimed and what I’d actually done. I started answering the way I’d talk to a colleague, not the way I’d write a resume bullet.

Example exchange that mattered:

Claude: “For each iteration of the consumer product you helped rebuild — what was your specific role? Were you defining the network protocols between device and backend? Choosing the OS and toolchain? Working on the device-side software?”

Me (paraphrased): “All of the above. One time the company we hired for the first iteration pushed hard to have us do all our transport over SCP because it was a baked-in library. I put my foot down and caused a big stir on this one. I said it was a showstopper, we couldn’t do it, it would not scale. Instead I pivoted them to use HTTPS, which back in 2007 was not as common. This allowed them to scale and was recognized as the right decision.”

Claude: “This is one of the best bullets in your entire career and it’s currently invisible. Pushing back on a contracted firm in 2007 to insist on HTTPS, when HTTPS-at-scale was uncommon, and being proven right — that’s exactly the kind of ‘I have judgment about systems that pays off years later’ story senior infrastructure hiring managers want to see.”

That bullet wasn’t on my resume. It had never been on my resume. I’d told the story at a few interviews over the years but had never thought of it as resume content because it was 18 years old and didn’t have a metric attached to it. This pattern repeated a couple dozen times across the session.

The single biggest unlock: mining old performance reviews

This is the part of the process I want you to actually do. Everything else in this post is technique. This is the move that changed the outcome.

About three hours in, I had a thought: I have nine years of performance reviews sitting in my Google Drive. Both the ones I wrote about myself and the ones my managers wrote about me. They were originally written for an internal audience, but they’re an honest, time-stamped, contemporaneous record of what I actually shipped, in much more detail than I would ever reconstruct from memory.

It’s hard to overstate how much you forget. Specific project names, the percentage of fleet you migrated, the number of services you brought under monitoring, the engineer you mentored who ended up promoted. These blur into a general sense of “I did a lot of stuff,” and the resume can easily reflect that vagueness. The reviews remember what you’ve forgotten.

So I downloaded every performance review and self-review I could find, dropped them into a folder, and shared the folder with Claude. We read through them in batches, organized by the resume section we were working on at the time.

Prompt that worked: “I just downloaded every performance review and self review from the last 9 years. Walk through them and surface back to me: (1) any specific accomplishments not yet captured in our drafts, (2) recurring themes that should become resume threads, (3) any direct quotes from manager feedback worth adapting.”

The results were a bit embarrassing, honestly. In one half I’d identified a global vending machine outage affecting 23% of Meta’s fleet, caused by an incomplete IPv4 turndown, and coordinated cross-regional remediation that contributed to a 10% reduction in stockouts. I had completely forgotten about this. In another, I’d been the on-site engineer for a rooftop network buildout that supported a public millimeter-wave bandwidth demo that ended up announced at F8 2017. I remembered the demo; I’d forgotten my role in it.

Real numbers surfaced too — things I’d hand-waved past in conversation but had quantified rigorously at the time. Service monitoring coverage grew from 49% to 93%. Migrated 44% of the fleet from Windows to Linux. Brought RLM and FlexLM license manager fleets from 0% to 31% cookbook coverage. Performance reviews are written by people who have to justify ratings, so they’re packed with the specific percentages and counts that make resume bullets land.

Even more useful: themes. Reading three years of reviews back to back, I could see patterns I never noticed living through them one half at a time. The phrase “shifts the team from operational/reactive to strategic engineering” showed up across multiple managers, multiple years. That wasn’t a thing I’d articulated about myself, it was something I’d been doing reflexively for years. It became a thread in the new resume’s summary. Same with “identifies problems outside formal scope and ships solutions through partnership with the right teams.” I’d been doing that my entire career. The reviews told me.

A few practical notes for doing this yourself:

• Get both perspectives. Self-reviews capture the what. Manager and peer reviews capture the outside view — the framing that sticks with people, and the strengths you’d be too modest to claim. You want both.
• Pull as many cycles as you can access, while you still can. Meta keeps performance reviews accessible to the employee. Other companies vary. Check what you can still get to before you need it.
• Don’t read them all yourself first. Let the model read them. You’ll get to a useful synthesis 10x faster than slogging through years of self-assessment prose.
• Watch out for sensitive content. Reviews can contain coworker names, internal project codenames, comp discussions, and leveling specifics. None of that belongs in the resume. Tell the model to filter, and confirm it does.
• Look for the things you cut from old resumes. A specific bullet I’d had in 2019 and removed because I “didn’t think it sounded impressive enough” turned out to be the strongest evidence of cross-org leadership in my entire career. Past-you’s instinct was wrong. Trust contemporaneous past-you over self-editing present-you.

If you remember nothing else: the performance reviews are the unlock. Everything else is technique. The reviews are content.

Hype man, not yes man

The most important thing I told Claude in the whole session wasn’t a question. It was an instruction about how to behave.

When it comes to resume writing, you don’t want a yes man behind you. The default mode of an LLM is to be agreeable and accommodating, and an agreeable AI will happily encode all of your existing biases, including the modesty problem most of us have about our own work and the impostor instinct that tells us not to claim things that are actually true.

What you want is a hype man who knows your strengths. Someone who’ll tell you when you’re underselling something that matters, push you to claim credit for work you’re trying to hand off to “the team,” and call out the moments where false modesty is costing you. And someone who’ll push back the other direction when you’re inflating something. When you’re about to put a bullet on the resume that won’t survive an interview, or when you’re claiming scope you didn’t actually have.

That two-sided pushback is what makes the dynamic work. Most of us don’t have someone in our lives who’ll do this for us honestly. Friends are too generous. Mentors are too busy. A peer review at work is too political. The model, if you set it up right, will just tell you.

Prompt that unlocked this dynamic: “I want you to be very critical about the inclusion of these though, and not just agreeable. I know that a lot of what I will share is ‘good work’ but that doesn’t mean that it will strike the right tone in a resume.”

You have to ask for this, and you have to keep reinforcing it. The default is helpful and accommodating, and helpful-and-accommodating won’t tell you your favorite bullet is mediocre.

I caught Claude doing the right thing in both directions across the session. When I wrote a paragraph saying I’d recently become more rusty technically but that AI tools were leveling the field, Claude flagged that this is a specific failure mode in interviews. Saying “AI levels the field” out loud lands as evidence of exactly the crutch hiring managers are watching for. We reframed the content.

Later, I was reluctant to include the side project I run for my wife’s music licensing company, and Claude pushed back the other direction. Without it, the resume had no recent evidence of me shipping production code, which made the “I’d consider IC roles” option look like wishful thinking. We added it.

Those are the two failure modes a hype man is supposed to catch. Overclaim where you’d embarrass yourself in an interview. Underclaim where you’re invisible on the page. Set the contract for both up front and your conversation gets dramatically more useful.

The “master resume” pattern

After several hours of work, I realized we’d built something I’d wanted for years but never managed to maintain: a master resume.

The idea isn’t new. Career coaches have been telling people to do this forever. You don’t have one resume, you have a content repository: every bullet you might ever use, organized by role. For each application, you select the right 5-7 bullets per role and rewrite the summary to match the target. The smart, dedicated people in your life have probably been doing some version of this for years. Most of the rest of us tried it once, found the maintenance tedious, and quietly stopped.

That gap (good advice, real friction, not enough discipline) is exactly the shape of problem that AI-native workflows are good at closing. The reason it never stuck for me before is that updating the master by hand was busywork. The reason it sticks now is that I have a tool that can read the master, read a job posting, and propose which bullets to keep, which to swap, and how to reframe the summary for that specific target. The discipline tax is gone. Oh, and it tickles the “use a computer to solve a problem” part of my brain that keeps me going in this industry.

The master we built is four-and-a-half pages long. I will never send it anywhere. It exists so that the tailored versions I send are each as good as they can be, and so that next time I update my resume I’m starting from a complete inventory of my career rather than from whatever fragments my memory hands me.

This is worth doing even if you’re not job hunting. The master is a personal artifact. It’s the thing you’ll be glad you have in two years when something changes and you don’t want to start from scratch again.

The part that’s directly relevant to right now

Meta told us this year that we were all expected to become AI-native. For most of us that started as code completion, evolved into agentic coding workflows, and is now showing up in places we didn’t expect. Resume writing is one of those places. It’s a task LLMs are unusually well-suited to: long, conversational, requires holding a lot of context at once, benefits enormously from a second opinion, and produces a tangible artifact at the end.

If you’re updating your resume this week, because of the layoffs or because you’re paying attention to what they mean, don’t paste it into Claude and ask for a rewrite. Ask for an interview. Tell it what’s not working. Pull your old performance reviews and let the model read them. Set the contract for a hype man, not a yes man. Build the master.

The hidden value of the whole exercise wasn’t the document. It was the inventory — a clearer sense of what my career actually looks like from the outside than I’ve had in years. The resume is a useful byproduct.

If you’ve been told you need to become AI-native and you’re not sure where to start, this is a small, low-stakes, immediately valuable project. The output is useful regardless of what happens next. And the skill of running this kind of conversation transfers directly to the harder work waiting for us on the other side of the reset.