Making it up, changing my mind, and learning to COPE with AI

The older you get, the more you realise that much of the time, people are making it up as they go along – stumbling forward on rules of thumb, mental models and “experience”.

How AI impacts ‘work’ is Exhibit A. By my count, we’re at iteration nine of theories on this. Each one airtight until the data disproves the hypothesis and makes it look naive.

Predicting this kind of thing has a long, bad record. When Pope Leo XIV released his AI encyclical this month, he took his name from Leo XIII, whose 1891 encyclical warned that the industrial revolution would grind workers down. Bill Gurley did the maths on the All-In Podcast: in the 134 years since, the average work week fell from over 60 hours to 34, real wages rose eight to ten times, child labour in the US went from 18% to zero, workplace deaths fell forty-fold, life expectancy rose 60%, and global poverty dropped from 75% of humanity to under 10%.

Anthropic’s Economic Index maps theoretical AI capability against observed usage by occupation. Theoretical coverage reaches 0.8 across management, business, finance and computing. Observed usage barely touches 0.3 in any category.
Everyone is still figuring it out.

I’m not above any of this. My own calls keep flipping. I used to think SaaS was finished. The more I actually build with AI, the more I think that was a dumb idea. Systems of record — the Salesforces, the ServiceNows — probably survive. Do you really want your “truth” living in a vibe-coded database, or would you pay a little more to keep it in a battle-tested system? The value-added services layer sitting above those platforms is still up for grabs, and per-seat pricing is going to change — so margin erosion risk is real, and some valuation compression is justified.

But I think SaaS has reached a place where the risk-reward looks good again. Salesforce in particular is on my shopping list.

For all that uncertainty, companies are acting fast. The recent spate of AI job cuts has also produced some shockingly bad corporate communications.

Matthew Prince, CEO of Cloudflare, cut more than 20% of his workforce and offered a taxonomy to justify it: employees are builders (who create products), sellers (who drive revenue), or measurers (finance, legal, compliance, middle management). AI makes builders more productive, won’t replace sellers, and measurers are most exposed. “The vast majority of those we laid off last week were measurers. We cut middle managers across the organisation because AI allows us to have more direct reports per manager while still measuring and mentoring our teams effectively.”

The framing is interesting — and directionally feels correct. The communication was terrible.

Meanwhile Zuckerberg laid off 8,000 people after reportedly installing recording software on the survivors’ machines to study their work and train AI on it. Some of those let go had built AI tools at company hackathons to make their own jobs more efficient.

Aaron Levie (CEO of Box) seems to have the sanest take on this: CEOs are “uniquely prone to AI psychosis because they’re sufficiently distant from the last mile of work that still has to happen.” They see the demo. They never see the deployment.

To me, these cuts mostly seem like “AI washing”.

The work multiplies

My actual experience is AI creates more work. Dan Shipper’s essay “After Automation” – went viral and makes this case well — AI commoditises the residue of human expertise, collapses the value of default output, and creates demand for what’s different.

David Sacks (former US crypto and AI czar) might be slightly biased, but he makes the point well: coding is AI’s breakout use case, and software engineer job postings are still up 15% year-over-year to a three-year high. The Yale Budget Lab studied three years of data and found no discernible labour-market disruption from AI. Unemployment sits at 4.3%, near what economists call full employment. The one job AI automates most is the one that’s growing.

GitHub’s own data confirms the acceleration. Commits reached 1.4 billion per month — up ~13x in three years. Pull requests and new repositories show the same inflection. All three went vertical from December 2025 when agentic development workflows arrived. That’s a boat load of code for humans to manage and architect.

I think what mostly decides your exposure is the kind of worker you choose to be, and Mark Cuban put it best: “there are two kinds of people, those who use AI to learn faster than they ever could, and those who use it to avoid learning at all”. It’s pretty clear to me that this is going to be true: AI is not taking your job. Someone using AI more effectively than you will.

Employers should be training workers in real AI engineering — building harnesses, composing skills, architecting context — and it’s on all of us to spend our own time honing those skills.

Which is a nice segue to…

Mental Models for AI builders

So you decide to take the plunge and start building your own tools. You’ll need a way to think about it, or the skills and agents and cron jobs pile up into a heap of competing instructions.

The human-AI sandwich model

Start with the model Every.to calls the “human-AI sandwich.”Kieran Klaassen / Every.to — “You’re the bread in the AI sandwich.” From Dan Shipper’s “After Automation” (Every.to, 2026). The model: human sets the frame and parameters, AI collapses the task, human reviews, corrects, and decides what ships. You frame the task. The AI collapses it. You judge and extend the result. The human is the bread on both ends. It works beautifully for one task, and falls apart the moment you have a hundred of them wired together.

Cope – a better mental model for architects

That’s the job COPE does: Capture, Organise, Process, Execute.COPE draws on David Allen’s GTD and Tiago Forte’s PARA. I’ve reframed both for the agentic world — the AI handles the bulk of each phase, the human sets the frame and judges the output. It’s the model I use to structure everything the last piece taught me the hard way. Also I like the name cope, because that’s what we’re all doing.

Each phase has one thing that isn’t obvious. Capture is about format, and most of it runs on its own: clippers, feeds and email processors converting your tweets, newsletters and emails into markdown the moment they land. Organise turns that raw information into knowledge — stamp each file with date, source, topic and links, so you can build a graph and retrieval that holds up. Process is where the work happens: you point domain-skilled agents at that knowledge — a research skill, an analysis skill — and they reason over it and write the result, still in markdown.

Execute is where it becomes the finished thing — a chart, a page, a deck. That used to mean PowerPoint and Word. Increasingly I render in HTML instead: lighter, and a language both humans and AI can read — and, it turns out, enjoy. Every chart in this essay is HTML, rendered this way — no screenshots. The through-line is that every stage stays readable to both of you: markdown until the end, then HTML.

The full build — capture to execute, tool by tool, and the stack I’ve grown on top of it (73 skills across seven families, named bstack because apparently that’s the kind of person I’ve become) — one for a future memo.

Building is fun

If you want to start tonight, build a skill-manager first: a skill whose only job is to build other skills. Claude Code has an open-source one here. When the skills start delivering what you wanted every time, it’s satisfying in a way that’s slightly addictive.