The Recalibration

I keep, pinned above my desk in Kraków, a printout of four job descriptions we wrote between January and March of 2022. Senior designer. Senior engineer. QA lead. Content and marketing lead. Each one is a tidy single page, each reads as if the work were stable, and not one of them describes, with any accuracy, what the four people doing those jobs today actually do. I want to spend the first half of this chapter on one of them — Max, our QA lead — because his role has changed more than any of the others, and because his change is the cleanest single illustration I have of what the rest of the studio has lived through. Then I will turn the chapter around and speak to you, by role, about what to do with the same change.

Max is our QA lead. In 2022 his week was a rotation of regression runs, exploratory testing, bug triage and writing the test plans the rest of the team treated as homework. He ran the on-release checklist. He trained the juniors in how to find the kind of bug a user would find at three in the morning. He was, in a studio culture that sometimes treated QA as an afterthought, the last pair of eyes on any release we shipped. He was reactive by design: developers wrote code, automated checks caught what they could, and Max caught what the automated checks missed. The value he created sat at the end of the pipeline, and you could measure it by what did not make it into production.

In 2026 our QA lead sits at the other end of the pipeline. What changed was not only the verb but the place in the pipeline where the verb happens. In 2022 he found bugs. In 2026 he writes the rules that decide what ships before a human ever sees it. I have spent longer staring at that sentence than I would like to admit, because for a year after the shift began I did not see it clearly. I thought his job was the same job, done with better tools. The job changed category.

The rules he writes are called, in our internal vocabulary, quality gates, and they are programs in every sense that matters. Each gate is a small authored artefact that decides, before the work reaches a reviewer, whether it is allowed to proceed: whether the auth model satisfies the conditions this engagement requires, whether the data model will survive the first hundred real users, whether the prompt embedded in the product returns something the founder would be comfortable reading on a Monday morning, whether a regression has reappeared. The gates run inside whichever agent the project has been routed to that month (Devin most often, occasionally others), and the fixes happen inside that agent, inside a VM, on a full-screen recording that lands in a Slack channel only the founder and the team read. A Linear issue opens itself, attaches the recording, references the failing test, cites the diff and closes itself on green. Our QA lead watches all of it happen, and occasionally tells the agent it is wrong. He is a QA lead the way a control-tower operator is a pilot.

He used to be measured by how many bugs he caught. He is measured by whether the system he has authored catches them without him.

That distinction is psychologically harder than it reads on the page. The satisfaction of the old job was the find: the moment the tester, on the second-to-last day of the sprint, produces the bug the developer did not. The new job has no such moment. The new job is the design of the apparatus inside which bugs either surface immediately or do not surface at all. It is, to use a word the chapter on Sculpting will take further, architectural. He is architecting the test suite he used to execute.

The tasks that collapsed for our QA lead are the ones he spent most of his time on in 2022: regression authoring, first-pass test-matrix production, the bug-to-ticket-to-engineer handover that used to eat half a week. Localisation QA, which once consumed the better part of a month for a German release, now takes a few days. His rule of thumb for what stays in human hands: agents get the describable, playbook-able tests; he keeps anything requiring the user’s point of view, the edge cases, and the seams between roles and states. On at least one engagement he still finds more problems by hand than the agent does, and harder ones. What became genuinely new is evaluating model outputs at scale, which borrows more from statistics than from software testing; auditing the agent fleet itself, catching the week Devin started closing tickets without properly reproducing; and contextual QA, reading a prototype with a specific user in mind and reporting what that user will do on their second attempt, which no agent we have yet tried can do.

What surprised him is how much of his week he spends not looking at bugs at all. He rules his kingdom by setting its laws and then watches, mostly on dashboards, mostly at the edges. What was hardest was letting go of the satisfaction of the find itself. He had spent a career being the last pair of eyes. He now spends it being the architect of a system that has many pairs of eyes, none of them his. The frame is the one The Compression introduced: judgement takes more of his week. The verb he spends his time on has moved from find to design.

Max’s is the cleanest example, but the pattern is the same in four disciplines. What collapsed in every role was the repetitive middle: boilerplate, first drafts, scaffolding, routine QA. What became more valuable sat at the judgement-heavy edges: taste, architecture, user context, editorial nerve. What is genuinely new sits in the thin strip where a model’s output meets the product’s surface. The World Economic Forum’s Future of Jobs Report for 2025 estimates that thirty-nine per cent of workers’ core skills will change by 2030 and that fifty-nine of every hundred workers will need reskilling or upskilling inside the same window.¹ Those numbers read, from where I sit, like a description of what already happened at Sliday between 2022 and now, across the 200+ projects the studio delivered in that period.

The question is whether you are changing at the same rate as your job. So here is the edit, by role.

The frame is the one The Discipline of Subtraction introduced. A prototype that arrives on a Monday morning is over-built; the studio’s job is to subtract the welded-on weapons, keep the judgement that got it driveable and harden the parts that will carry users. The same three verbs apply to a career. Stop. Keep. Start. Subtract the tasks that have become free. Keep and harden the judgement that has become scarcer. Add the one skill every role in 2026 requires: instructing a machine well and reading its output critically.

I am going to speak to you directly, by role, because the advice cuts differently depending on the seat you sit in. If you are at a desk I have not named (and if you are in QA, Max above is your edit), most of what follows will still apply; the verbs will need translating into your own practice. Read the role closest to yours first, then read one you would not have thought to read, because the shape of the edit is more visible at a distance than up close.

If you are a designer, the collapse is the production: the initial screen, the component, the first cut of an empty state from a verbal brief. The model does those faster than you, and the founder will produce them on a Saturday anyway. What is now valuable is the cut — looking at forty screens and telling a founder, in order of severity, which twenty-two to delete — plus the client conversation and the final-mile craft: the sidebar that will not sit, the form that flinches under a long Finnish word. What is new is the prompt repository: move your component library into code, write your tokens in a format a model can read, version the prompts, and open the repository before you open Figma. The prompts are, literally, a program you are writing, and the quality of the program determines the quality of the screens that come out. This week: pick one screen-production task you still do by hand and move it into a prompt in the next fortnight. If you cannot, tell the person who manages you; that is a conversation about tooling or training, not a failure on your part.

I should qualify that, because two of our own designers answered the same question a fortnight apart and answered it in opposite directions. Valeria’s week has inverted: most of her hours now go on editing what a model produced and on briefing it, where two years ago they went on original decisions. Artem’s week, our other designer’s, has not. He still spends about seven in every ten hours making original design decisions, against a 2024 week that was almost entirely manual, and he uses the model to organise his tasks and check his output rather than to draw. Same studio, same tools, same two years. The edit does not land the same way on every desk, and a designer whose split has not moved is not therefore behind.

If you are a developer, the collapse is boilerplate, scaffolding, the first pass of a test suite, the first draft of a migration, the first cut of a pull-request description — and the polite half-hour at the start of a ticket where you worked out what it was asking for; the model does that now, with uneven results, useful on some easy tickets and wrong on harder ones, and telling the difference is a different skill. What is valuable is reading code in volume with taste, the deletion branch where your value is in what you remove, the hardening work (the stubbed auth, the hard-coded secret, the integration without rate limiting, the data model that will not survive its first real user), and architectural judgement. Most of it is unglamorous; all of it is what the engagement now pays for. One senior engineer put it plainly: be the person who can own a problem end-to-end, from business framing to launch — the market is about to be flooded with people who can ship code fast and starved of engineers with the range to do the rest. What is new is prompt reviews, evaluations for model behaviour inside the product, pairing with an agent on work whose commits and mistakes are still in your name, and a daily fifteen minutes on one newsletter deep enough that you are not surprised by what your tools did overnight. This week: count the lines you added last month against the lines you deleted. If the ratio has not shifted towards deletion since 2023, your work has not shifted enough. Show the ratio to your tech lead. Have the conversation.

Those two roles carry most of the Sliday payroll. The other three matter no less; what follows is shorter by design.

If you are a data analyst, the collapse is the first pass — the SQL, the exploratory chart, the cleaned CSV, explaining what a quartile is in a meeting. What is valuable is the part that never showed up on your CV: asking the real question behind the stated one, knowing which metric will actually move the decision, and being able to tell a CFO this is not the number you should be looking at, and here is the one you should. What is new is the evaluation layer — you are increasingly the person whose test suite decides whether the model’s analysis is trustworthy, auditing the dashboard the model generated overnight. Your work moved one layer up: from producing the number to certifying it. This week: pick one dashboard you maintain by hand, write a prompt that produces its first draft, run the two side by side for a fortnight. If the prompt is within eighty-five per cent of your output, start negotiating what you do with the time you will get back.

If you are a content manager, the collapse is drafting, rewriting for length, A/B variants, keyword-optimising a page — at volume, in a style within a fortnight’s editing of sounding like you. What is valuable is brand voice defended line by line against a model’s pull towards pleasant mush, the commissioning brief that tells the model what a piece is for and what it must not sound like, and the editorial nerve to cut a paragraph the model is proud of. What is new is the voice repository: collect the sentences that sound most like you, annotate what makes them yours, feed them to the model as reference and version the voice document the way an engineer versions a schema. The bottleneck stops being your typing speed and becomes the volume of genuinely good ideas the studio has to say — harder to fix, and more honest. This week: pull ten posts from your archive that sound most like the brand, write one sentence each on what makes them yours, paste all twenty into the system prompt you use for first drafts, and judge what comes back on Wednesday.

If you are a project manager, the collapse is most of what a status update used to be — writing it, collating it, nagging the team for bullet points; the model reads the repository, the Linear board, the Slack thread and the calendar and produces a first draft closer to the truth than half the reports I received in 2022. What is valuable is what a good PM always did and the profession undersold: holding the reality of the engagement in your head across client and team, scope discipline, the hard Thursday meeting where Tuesday’s request costs Monday’s, and narrative across time — this is what we said in week one, this is what we learned in week five, this is why we now recommend something different. What is new is running the model fleet: you are, whether your job description says so or not, a junior operations person for a small team of agents — watching their cost line, noticing when Devin closed tickets without reproducing, flagging when the auto-generated Slack post is lying sweetly about what was fixed. The honest unit of management has widened from team of humans to humans plus agents plus the glue between them. This week: write down the five agents your engagement runs and what each cost in the last seven days. If nobody has that ledger, start it.

The new skill underneath all five roles has a name now, though nobody quite agrees on it. Andrej Karpathy, in his Software 3.0 talk at Y Combinator, offered the phrase that gets quoted most: everyone is a programmer.² It is the most misunderstood slogan of this era. It does not mean everyone writes code; the designer does not, in any sense her 2022 self would have recognised. It means everyone now programs something — an agent, a pipeline, a prompt chain, a document that instructs a model — and is paid for the quality of that programming. A designer’s prompt repository, a content manager’s commissioning brief, a QA lead’s evaluation suite, a PM’s agent-ledger: each is a program. The line between code and not-code has moved, and the new line runs through every desk in the studio.

The hardest thing about this advice is that the old task list — the one you are being asked to subtract from — is exactly the list that got you hired, praised and promoted. You are good at it. You have, in most cases, spent the best decade of your working life getting good at it. Subtracting from it feels, in the body, like throwing away a trophy. The people who do the edit early pay six months of professional discomfort and are repaid within eighteen.

I say this as someone who has had to do the same edit twice. In the first half of the 2010s I made money extracting specifications from founders’ heads; that work is now gone, and I did not see it go fast enough the first time, which cost the studio a quarter we should not have lost.

The bill for this falls hardest in two places, and they are mirror images. The Stanford Digital Economy Lab’s August 2025 working paper, with the deliberately quiet title Canaries in the Coal Mine?, found that employment for software developers aged twenty-two to twenty-five dropped by roughly twenty per cent between 2022 and 2025 — in the same window the roles above were being reshaped upwards.³ The juniors who would, in the old shape of the industry, have spent a year writing the boilerplate those seniors no longer write are finding fewer entry-level seats to sit in. This is not a student problem; it is a studio problem. The tasks we used to give juniors are precisely the tasks the model now does, and those tasks were how a junior became a mid-weight engineer or designer or marketer. We tried one short apprenticeship window ourselves, and most of what we had planned to teach had been automated out from under the curriculum before we began. If nobody at the studios currently hiring figures out a new apprenticeship — what a junior is paid to do, what they learn by doing it, how they graduate out of it — the pipeline that produced the seniors of 2026 will not produce the seniors of 2036.

The other place the bill falls is on the mid-to-senior people whose skill used to distinguish them from a junior. Clinging to that task list is the fastest route to being laid off in 2026, and the hardest to see coming, because nothing visibly goes wrong for the first twelve months. Your hours stay steady. Your reviews stay warm. Your work is still good on its own terms. The slope underneath you is the engagement mix changing around you, and you will not see it until the old engagements stop arriving. By then the edit is late.

So do it now. Write the one page. What I stop doing, what I keep, what I start. Re-read it in six months. If the page has not changed at all, the page is wrong; if it has changed entirely, it is also wrong. Something in the middle is the work.

Every career that survives a methodology shift survives by doing to itself what the methodology does to the work.