Annie Tsai

Documentation Is the Critical Career Skill in the AI Era

I’ve run a lot of workshops over the years helping people audit their portfolio of work and stories to be able to clearly articulate their “special sauce,”  and this is both an evolving story as one’s body of work builds and one that far too many people don’t think about until it’s too late.

Most of the people in those rooms have a specific, valuable thing they do that others can’t replicate, a way of working or a set of outcomes they’ve produced or a combination of skills applied to a particular kind of problem. What they almost universally don’t have is a clear way of describing that thing, and even fewer have the documented record to back it up when it counts, because they’ve been too busy doing the work to track the work.

This has always been a gap, but the AI era has made it a critical one. AI can screen a resume in microseconds, surface fifty candidates who look similar on paper, and generate an interview answer that sounds both competent and forgettable in equal measure, and the only thing that consistently cuts through that is specific, documented proof. “I improved the onboarding experience” is going to get tossed to the side, but “I redesigned the onboarding flow and 7-day activation went from 34% to 51% across 12,000 new users a month” gets you a conversation. It is specific outcomes with numbers attached that become the differentiator.

And yet, too few ever take the time to log this information before it’s forgotten.

The 2026 context makes this an urgent habit to build

Layoffs.fyi is tracking 165,269 tech layoffs across 1,064 companies in 2026 alone, and Challenger, Gray & Christmas reported AI as the stated reason for 26% of all job cuts in April — the second consecutive month AI led the list. The roles being cut hardest are also the roles where business impact is hardest to quantify and least likely to have been written down: customer support, quality assurance, entry-level engineering, HR coordination. Brynjolfsson and his colleagues at Stanford’s Digital Economy Lab, working from ADP payroll data covering millions of workers, found that early-career workers ages 22 to 25 in AI-exposed occupations have seen a 16% relative decline in employment since 2022, and it’s showing up as a hiring freeze rather than a mass layoff event, which means it doesn’t register in the trackers but is very real to anyone trying to start a career right now.

Most of those people — capable, with real impact behind them — are now trying to explain to recruiters what they did and what it was worth. What comes back is usually something like: “I worked on the product team and was involved in improving the onboarding experience.” The number lived in a dashboard right after ship, then the quarter ended, priorities shifted, and the metric got buried.

Why memory alone doesn’t work

Hermann Ebbinghaus established in 1885 that human memory decays exponentially — we lose roughly 56% of new information within an hour and 79% within a month. The curve doesn’t care how significant the work was or how much it mattered to the business, and it runs the same math on everyone.

The forgetting curve isn’t the only force working against you, either. Christine Exley and Judd Kessler at NBER documented what they called the gender self-promotion gap: on identical performance, women consistently rate their own contributions lower than men — 46 versus 61 on a 100-point scale on equal output. Memory is reconstructive, not a replay; we rebuild from what we can access, and what we can access is shaped by what we wrote down. If your only source for “what was that project worth” is your own unaided recall, months after the fact, and you’ve been socialized to underestimate your contribution to begin with, the self-promotion gap shows up looking like a mindset problem when it’s actually a documentation problem.

Correll and Simard’s research found a parallel pattern in performance reviews: women tend to receive vague evaluations (“you had a great year”) while men tend to receive specific, metric-grounded feedback — a growth number, a client win, a conversion improvement. When you’ve been given the number, you know the number. When you’ve been given an impression, you’re reconstructing from an impression, and impressions compress faster than data. In this world, the most reliable fix for a documentation gap is documentation: a file, updated when you have the number, rather than mindset coaching after the fact.

Logging at the moment of knowing

The advice to “track your accomplishments” is both right and too abstract to become habitual. What breaks down is timing and friction. By the time a monthly accomplishment review rolls around, you’re working from memory — the lossy format Ebbinghaus already described.

What actually works here is logging at the moment of knowing: right after the ship, right after the dashboard updates, right after the A/B test concludes, right when you have the outcomes data to show the results. The before-and-after is what makes a resume bullet real. “I redesigned the onboarding flow” is a task description. “I redesigned it and activation went from 34% to 51%” is an impact statement. That delta, logged when you have it, is the difference between a story you can tell and a vague gesture at work you did.

This week’s gift: Work Receipts

I built a free, open-source Claude skill for this: Work Receipts.

It runs as a slash command in Claude Code (`/wr`) or conversationally in the Claude app — no installation required if you already use Claude. You describe what you shipped. It asks the questions you’d otherwise forget to ask yourself: “What was it before? How many users were affected? What was your specific role, not the team’s?” If you don’t have the number yet because the test is still running or the dashboard hasn’t updated, it logs the intent and resurfaces the question later. The output formats on demand: resume bullets, STAR interview stories, a LinkedIn narrative.

What I love about the Work Receipts skill is that over time, it’ll look at the body of your accomplishments and start building a narrative for you that perhaps you didn’t see in your own work. Perhaps it’ll find that you are particularly skilled at fixing incredibly complex user interface design flaws. Or maybe you happen to always be the one brought in to de-crisis a crisis situation across multiple vectors. Or, maybe you saw a 2% lift but in 60 days of daily Work Receipts updates, the data actually shows that the project you delivered yielded a 12% lift. You not only now have the ability to track accomplishments more accurately, but will also be able to improve your storytelling.

I built it after noticing the pattern across years of conversations with the same pattern, and after nearly letting my own numbers slip past me the same way. I’d been telling myself I’d remember, but too often I forgot too.

Your Takeaway This Week

Start tracking your wins, outcomes, key metrics, learnings — track it all. No one else is going to do this for you, and in the AI era, this may be the single most important thing you do.

Log one thing today. Open Claude and send it this prompt:

“Review this github repo and interview me for the answers needed to personalize it for my Claude instance. install the skill: https://github.com/meannie/MomsinTech/tree/main/WorkReceipts”

You won’t need to give it much information. Then, describe the last meaningful thing you shipped.

Get specific with the baseline. The most commonly missing field is the before. What was the error rate before you fixed it? What was the activation rate before you redesigned the flow? The delta is what makes a bullet compelling, and the baseline is what makes the delta real.

If you’re already in a search, your metrics still exist somewhere in your old company’s systems — export what you can before the access window closes, and use Work Receipts to shape them into something specific enough to say in a room.

And if you’re a leader, consider making this a team practice. Your people’s impact becomes visible to them rather than only to your dashboards, performance reviews get real data rather than reconstructed impressions, and the people on your team who leave will walk out with something real to show for what they built.