How to Rebuild a Job Description for the AI Era Using Coinbase's Interview Playbook

Use Coinbase's published interview rebuild to rewrite one JD and one interview loop today, with any chatbot.

So back in the mid-90s I went to a job fair looking for a finance job. While I was standing around, somebody at a booth said the magic words: "We're hiring Internet service reps. No experience required." I'm like, free training? I'll do that. Four years later I'd gone from hanging up on John Sculley because I didn't know who he was — to running a 2,500-person, hundred-million-dollar business unit at 29.

Here's the thing about that job posting: it was wrong the day it was printed. "Internet service rep" described a job that reinvented itself every six months I held it, and the paperwork never caught up. And most job descriptions I've read in the thirty years since have the same problem. They describe how the work used to get done, written by somebody who did it the old way.

That gap is the widest I've ever seen it right now, because AI changed the work faster than anyone's HR templates. So this week's Lesson is one move: take one job description on your team and rebuild it — plus one interview exercise — for how the work actually gets done with AI. Coinbase just published a year of homework so you don't have to guess.

// The Takeaway: Your job descriptions describe how work got done before AI. Today, pick one role you manage and rebuild its JD — and one interview exercise — around directing AI, evaluating its output, and catching its errors. Coinbase spent a year rebuilding its engineering interviews exactly this way and published the results. The prompt below turns their playbook into about an hour of work with any chatbot.

Read Coinbase's interview rebuild — it's a ten-minute read → · About an hour total. Free tier of ChatGPT, Claude, or Gemini works. No new tools — just one current job description and ten minutes with the person who holds the role.

Why Coinbase? Because they didn't write a think piece — they published receipts. AI-generated code went from 5.7% of their codebase in Q1 2025 to roughly 80% by Q1 2026, all human-reviewed. Then they noticed their interviews were still testing whether candidates could work like it was 2023. What they changed, and how carefully they changed it, is the best public template I've seen for any team — not just engineering.

// The real shift: The scarce skill is no longer producing the work. It's directing the work and catching the errors. As one Coinbase leader put it: "When the cost of building goes to zero, the cost of identifying what to build, verifying it's correct, and getting it out safely becomes the limiting factor." A job description built around production verbs — writes, builds, produces — is screening for the cheap thing and ignoring the expensive one.

The short version of their post, because the specifics are the lesson:

Their old interviews failed in both directions. False positives: candidates who memorized patterns but lacked judgment. False negatives: strong architects who stumbled on recall trivia. And when they analyzed the data, two of their interview rounds correlated at 84% — they were paying twice for the same signal, and neither round measured how a candidate used AI at all.

The new interviews turn the AI on. Candidates work in realistic, existing codebases with AI tools fully enabled, and the interviewer scores prompt quality, how the candidate evaluates the output, and whether they catch the errors the model introduces confidently. Judgment over typing.

They screen for three dimensions of AI fluency, weighted the same for juniors and seniors: usage (picking and applying the right tool), application (knowing when AI is and isn't appropriate), and understanding limits (spotting where AI breaks, including privacy and security).

They kept discipline. No added interview rounds — every new signal had to replace an existing round or measure something distinctly different. They piloted on frontend roles in late 2025, expanded to backend in January 2026, and went company-wide in March 2026. Pilots with exit criteria, not a memo from on high.

Swap "codebase" for contract, campaign, forecast, or close process and none of this is engineering-specific. Which brings us to the part you can do this afternoon.

Here's the prompt. Copy it whole, then fill in the two brackets:

One honest caveat, because Coinbase was honest about it too: their evidence is early. Candidates who pass the AI-assisted assessment advance at a meaningfully higher rate, but the sample is small, and they flag a "half-life problem" — models improve fast enough that a good exercise goes stale. Their answer is quarterly review. Yours is simpler: re-run this hour each time the role reopens.

Paste your own job description into the same prompt. I did it with mine and the audit was, let's say, humbling — a decent chunk of what I told myself was judgment turned out to be bucket (b), reviewing things a model drafted. That's not bad news. It's just the first accurate org chart entry I've produced in years.

This is the same week 26 Meta employees sued over AI-driven layoff selection — how companies use AI on people decisions just became contested ground, and yesterday's Tangle covered the whole front. Hiring is the one people decision you get to rebuild proactively, on your own terms, before it rebuilds you. The job description is the smallest unit of workforce strategy. Rebuild one today, and the next person you hire will be hired for the job that actually exists.

But anyhow — start with one JD. The one you'd have to post next.