Build an AI-era portfolio that does the talking

Somewhere in the past two years, "experience with AI tools" stopped meaning anything on a resume. Everyone claims it, so the people doing the hiring quietly discount it — and go looking for proof instead. The good news: proof is buildable. A small portfolio of three honest artifacts will do more for your next role than any rewritten bullet point, and you can assemble most of it from work you've already done.

Why claims stopped working

Put yourself on the other side of the desk. A hiring manager reads "leveraged AI to accelerate delivery" on the fourth resume in a row. What have they learned? Nothing they can act on. The phrase covers everything from "typed a prompt once" to "shipped an agent that three teams depend on," and the reader has no way to tell which one you are. When a signal becomes free to claim, it stops being a signal.

What's scarce now — and what interviews are increasingly built to detect — is evidence of judgment: what you built, why you built it that way, and how you knew it worked. That evidence is hard to fake and easy to recognize. Which makes it exactly the thing worth spending a few weekends producing.

The three artifacts that carry weight

• A shipped workflow or agent, end to end. Small is fine. A working tool with real users — even three of them — beats an ambitious repo of scaffolding.
• An eval set with results. The before/after table that shows you can define "good," measure it, and improve against it.
• A failure write-up. One page on something that broke, why it broke, and what you changed. Nothing signals seniority faster.

The first artifact proves you can finish. It doesn't need to be a product — a pipeline that triages your team's support tickets, an agent that drafts your weekly report, a retrieval tool over your own notes. What matters is that it runs, that someone besides you has used it, and that you can show the design decisions you made along the way.

The second proves you can tell whether it's any good. Evaluation is the skill hiring managers say they can't find, and it photographs well: a table of twenty cases, what the system got wrong before, what it gets right now, and the two cases that still fail. That single screenshot starts better interview conversations than any architecture diagram.

The third proves you have judgment under failure — the thing AI systems demand constantly and resumes never show. Write up the day the model started confidently inventing customer names, or the eval that passed while users complained. Name the cause, the fix, and what you'd instrument differently. Engineers who can narrate a failure calmly read as people who've actually operated something.

Write the decision record, not the README

What separates a portfolio piece from a code dump is the narrative that rides along with it. For each artifact, write five hundred words answering four questions: what constraint were you actually facing, what options did you consider, what did you choose and why, and what would you do differently now. That document — not the code — is what a busy reviewer reads, and it's where your thinking becomes visible.

Resist the urge to polish the story into a straight line. "We tried the obvious approach, here's the specific way it failed, here's what we did about it" is more credible than a flawless victory lap — and far more memorable.

Package it for a ten-minute read

Assume your reviewer gives you ten minutes, generously. One page, three artifacts, each with a two-sentence summary and a link to the fuller write-up. Lead with the artifact that matches the role: the eval work for platform teams, the shipped agent for product teams, the failure analysis for anything with "senior" or "staff" in the title. The page itself is a work sample — if it's clear, scoped, and honest, you've already demonstrated the job.

A weekend-sized start

You don't need new projects — you need to excavate the proof from work you've already done. This weekend, pick the one project you could defend in front of a whiteboard, and write its decision record. Next weekend, rebuild its eval set well enough to screenshot. The weekend after, write up your best failure. Three weekends, three artifacts, and your next application opens with evidence while everyone else's opens with adjectives.

The market hasn't stopped hiring people who work with AI. It's stopped believing them. Be the candidate who doesn't need to be believed — just read.