The Workshop
A record of what we have taken on. Each entry started as something that could work better, and each one is written honestly: what the real problem was, what it took to solve, and what it actually produced. Including the parts that did not work.
Adoption is not the same as capability
I was asked to drive Microsoft Copilot adoption across an organization of ninety people. I did the things you do. Videos. A newsletter. Sessions, guides, prompts, examples.
Adoption moved, but it wavered. And the people who did pick it up were using it to save a few minutes on the obvious tasks. Summarize this. Clean that up. Useful, but nowhere near the ceiling.
As a heavy user, I knew what was actually available. The gap was not enthusiasm or access. It was judgment. Nobody had been taught what to delegate to the tool and, more importantly, what to keep. Without that line, people either underuse it or hand over work they should have held.
I focused on leaders deliberately. If the people running teams developed real command of the tool, it would move through their organizations far more effectively than another training deck ever would.
So I wrote the book to give them that: a working system for using AI without surrendering the decisions that are theirs to make. The specific tool will change. The judgment transfers.
The advice that gets you started is not the advice that keeps you going
Assign a role. Give it context. State the format. That is sound guidance, and it is in the book, because it is where everyone should start.
I rarely do any of it now.
Across a working stack of tools, and a lot of building, the way I prompt has changed. The role-and-context scaffolding has become situational: something I reach for when I need to force a specific direction, not a default I apply to every request. What replaced it is a set of instructions I return to constantly because they reliably produce better output.
So I wrote them down, along with the reasoning behind each one. Not a list of prompts. Thirty-two blocks, each with what it is for, why it works, what it has caught, and when not to use it.
The prompts matter less than the reasoning. A prompt you copy works once. A prompt you understand, you can adapt to the next thing. It grows as the practice does.
The gap between knowing and knowing precisely
Revenue is what the business is measured on. Bookings are what most teams are compensated on. Those are not the same number, and the distance between them is where leaders get caught.
You can beat a bookings target decisively and still miss revenue by a wide margin. Every leader knows this in principle. Far fewer can walk through exactly how it happens, or explain it fluently to a finance partner who lives in it every day.
That was not a knowledge gap so much as a precision gap. The understanding was there. The command of it was not, and command is what lets you make the right call under pressure and defend it to people who know the numbers cold.
So we built the training to close it. Ten modules, five working simulators, quizzes, and a full glossary. Not an overview. A working fluency in the number the business is actually judged on.
Nobody is reading carefully. Write so it survives that.
I finished my degree during Covid, going back to close something I had left open. Since then I have never really stopped looking. Not out of desperation, out of the habit of knowing what my options are.
Every year or so the search got serious again, and every year or so I hit the same wall. The resources were scattered across thirty tabs that never talked to each other. And I could not translate my own work into the language the process wanted. I knew what I had done. I could not say what it was worth.
So I built what I needed. An agent to work the resume itself. A scanner to check it against a job description before sending. A calculator to turn a metric that moved into a number that meant something. I used it. It produced interviews. I stayed where I was, for reasons that had nothing to do with the tools.
Then it sat. I put it up, sent it to a few friends who were searching, and never pushed it further. The build worked and the distribution never happened. That is the honest end of this one.
What I thought the problem was
I thought the problem was the machine. That is what everyone tells you. Applicant tracking systems eat your resume, and unless you feed them the right keywords, no human ever sees it. The number that gets quoted is that 75 percent of resumes are never read by a person.
I built against that assumption. Then I went and checked it, because a claim that convenient to the people selling the solution deserves a second look.
It does not hold up. The recruiter Jan Tegze went looking for the study behind it and could not find one. The figure traces to marketing material from a company that no longer exists, repeated in a Forbes piece written by the founder of a resume service, then cited for a decade by people who never checked. There is no study.
When someone finally asked recruiters directly, the answer came back the other way. In a 2025 study, 92 percent said their systems do not automatically reject resumes for formatting, missing keywords, or a low match score. Auto-rejection was used by 8 percent, almost entirely for hard compliance knockouts. Do you have the license. Are you authorized to work here. The kind of question where the answer is genuinely disqualifying.
So the robot is not throwing your resume away. That story is sold by the people who profit from your believing it, and I nearly built on top of it.
What the problem actually is
The real thing is harder to fix and less satisfying to blame.
Harvard Business School and Accenture published a study called Hidden Workers: Untapped Talent. It found large numbers of qualified people being screened out of hiring processes, and it put the cause somewhere specific. Not the algorithm. The configuration. Job descriptions inflated with requirements nobody actually needs. Filters that penalize employment gaps, non-linear careers, and unconventional titles. Rules written by people, encoded into software, and then left alone.
The system prefers careers that look tidy on paper. That preference was a human decision. The software only enforces it at scale.
Now add volume. I cannot give you a clean number here and I am not going to pretend otherwise, because the figures in circulation come from vendor blogs and expert estimates rather than published data. But the direction is not in dispute. Applications per posting have climbed steeply since 2022, and a large and growing share of them are written by AI. Fluent, competent, and indistinguishable from each other.
So the person reading yours is reading it fast, late in the day, against a stack of documents that all sound the same, holding a requirements list somebody over-specified before you ever arrived.
What that changes
Once you see it that way, the fix stops being about keywords and starts being about translation. There are two translations happening, and most people fail both.
The first is between your experience and the requirement as written. Not stuffing the words in. Making the substantive connection visible fast enough that a skimming reader does not have to construct it themselves. If a requirement says "renewal forecasting" and your resume says "led the annual planning cycle," you have done the work and you have not made the connection. You are relying on the reader to build the bridge, and the reader does not have time.
Keyword alignment matters here, but not as a cheat. It matters because it is a proxy for substantive alignment. Get the substance right and the language follows. Get the language right without the substance and you have talked your way into a room where you will be found out in twenty minutes, which wastes your day and theirs.
The second translation is harder, and it is the one that stopped me cold.
I knew my work had mattered. I could not say what it was worth.
Every operator has written this bullet. "Improved efficiency." "Reduced churn." "Streamlined the process." All true. All dead on the page. Because a metric moving is not a result. It is an input. Churn down two points is meaningless until you know how many customers, at what value, over what period. Two points on a small book is a rounding error. Two points on a large one is somebody's quarter.
The gap is that the person who did the work almost never has the arithmetic in front of them. You know the before and the after. You do not know what the delta was worth in the only unit the business actually reads. That is not a resume problem. That is a translation problem that happens to show up on resumes.
What got built
Three tools, one for each failure.
Works through structure and substance, not formatting tricks. The job is to make the work legible fast.
Takes a job description and a resume, and surfaces where the requirement and the experience do not connect. Not to help you game a filter. To show you where a skimming reader would have to do work you should have done for them.
Takes a metric, a baseline, a new value, a volume, and a value per unit, and returns what the movement was actually worth. It exists because I could not do that arithmetic in my head and neither can anyone else.
Then the rest of the stack, because I needed it and it did not exist in one place. A tracker so I stopped losing threads across applications. Templates. A qualification calculator for the honest read on whether a role was worth the effort. And the library of resources I had already spent weeks assembling. None of it was a product. It was one person's working set.
What happened
I used it and it worked. The resumes it produced got me in front of people. I had real conversations, and I got far enough into several of them to make an actual decision rather than an aspirational one. I stayed where I was. That was the right call, for reasons that had nothing to do with the tools.
Then it sat. I put it online, gave it to a few friends and colleagues who were searching, and they used it, and it helped them. I never pushed it anywhere. It has been a private thing I hand to people in my industry when they need it.
The build was fine. The distribution never happened. Naming that is more useful than pretending the ending was cleaner than it was.
What comes next, honestly
The tools were built against models that have been superseded more than once since. An alignment scanner written against what was available then is not the alignment scanner I would build today, and shipping it because it exists would violate the only standard that matters here.
So they get rebuilt. That work is on the bench, and it will be documented as it goes, including the parts that break.
What does not change is the thesis. Nobody is reading carefully. That is not a complaint about hiring, it is a description of the conditions. Write so your work survives it.
Sources
- Hidden Workers: Untapped Talent, Harvard Business School and Accenture, 2021.The spine. Attributes exclusion to process design, not to autonomous algorithms.
- Jan Tegze on the origin of the 75 percent claim, 2022.Traces the figure to marketing from a defunct company. Practitioner analysis, and the closest thing to a study anyone has produced.
- Enhancv recruiter study on automated rejection, 2025.Sample of 25 recruiters. Small. Directionally strong and the only direct evidence on the question, so it is cited with its size stated.
- Figures on application volume and the AI-generated share of applications.Every number in circulation traces to a vendor blog or an expert estimate. The trend is real and the numbers are not defensible, so the claim is made without one.
A calmer companion
Still in testing and still being refined. The full case gets written when the thing is real enough to stand behind.
Three agents that run the working day
Running here every day. Not yet packaged for anyone else to run, which is a different problem from building it in the first place.
The current problem is being worked
Cases start here, unsolved, before we know how they end. When one lands it moves up the page.
Follow along →