In the Anthropic webinar Jean and I did, I quipped that I feel like "10 Claudes in a trenchcoat." A good line for that audience, but the reality varies. Sometimes it feels so empowering: look how much I'm getting done! Look at what's possible now! And sometimes it feels overwhelming and unstable. It's not a neat stack. It's a tornado.

How often has that been the case, though? How often did I used to feel like four directors in a trenchcoat? Every summer when I was vacation cover for people below me, for sure. Every time I was covering a hole. More than that, probably, because that way of working becomes a habit that's hard to extract yourself from.

I was interviewing for another role once, with a VP who gave me the impression he was more four directors in a trenchcoat than a VP, which is probably why we were talking. Still, I don't recommend asking that question directly (oops). That impression was also part of what made me think I wouldn't learn from working with him, and I pulled out of the process.

As with many things in leadership, it comes down to what is managing up and what is managing down. The higher you rise, the more managing up is the job. But part of being a good leader – and even more, feeling like a good leader – is managing down. How much, then, do you end up filling the gaps below you, and how hard is it to do that well? How much are you in details that should be abstracted, checking on things where setting direction should have been enough?

The thing I miss most about my last job is the layer underneath me. Not the headcount or the intensity of my calendar, but the feeling of team and mutual reliability. When something went wrong, it didn't land on one person. I could ask people to handle things. I could find out later that things had been handled without me knowing. I could take time off and things would keep moving. We could absorb shocks because the weight was shared, and it went both ways: if anything happened, I was almost always the first ping.

The downside was that sometimes I felt flattened against a glass ceiling, with the fear – even guilt – that it wasn't just holding me down, it was holding down the people underneath me as well. Compressed between the two, with no idea how to advance what was fair and right.

Living without that limit has also meant living without the baseline. There's no ceiling on what I can do. There's also very little turning to others and trusting they've got it.

This is the output/responsibility gap. Ten Claudes give me a lot of capacity to get things done. They don't take any of the responsibility. That's all still mine.

If Claude had feelings, I'd worry about offending it, because it is relentlessly keen to help. Multiple times a day it tells me "I can draft that for you." And sometimes, yes. The email to the insurance company telling them not to automatically charge me for a product I don't want does not deserve my attention. The email to a friend absolutely does.

For me it comes down to responsibility: what can be delegated, what cannot, and what sits in between.

Delegated: coding, pulling data. Things I can describe, send away, and check when they come back. It's an input into something else.

Pairing: writing. Specs, blog posts, this newsletter, social posts. Always my take, only things I've actually read or actually think. It's good at breaking through the inertia, pushing on the structure, telling me where a paragraph isn't landing. Feature rollouts, too: it does the automated checking and I own the gates.

Human only: coaching and DRI feedback. Otherwise, what's the point?

For DRI, we are so AI-leveraged everywhere we can be: the course platform, data and checks, spell check and consistency passes. But the feedback is personal. We read everything, and we respond as humans. The leverage lowers the price point we can do it at. The human time is what sets the floor.

Which makes me wonder about the AI-only company. What's the point of it? What does it build? Who does it build for? The Sims, but make it programming – my constant refrain about some of the most AI-pilled ideas around.

I had a rough week recently. I'd committed to an October feature, and the scope grew: additional requests, and things I'd underestimated (Stripe, mainly). So it became a Friday release (I know, bad) that slipped into Saturday (I know, worse). Then I was functionally on call all weekend, phone on loud and ready to jump on any alert. Thankfully, undramatic.

On Friday night Claude told me it was time to stop working roughly every hour from 4pm until 10pm – when I actually finished. I ignored it. I had committed to shipping in October and damn it, the thing was going out.

AI can work all weekend. It cannot feel responsibility for a human-to-human commitment to deliver something. That was mine.

Heading into 2027, two predictions.

One: a leadership reckoning. Meta drafted thousands of engineers into an AI training unit this spring and walked it back by June, with Zuckerberg admitting mistakes in a memo. Shopify's CTO, hired for his AI expertise, quietly moved to an advisor role in October, a week after appearing in a Codex ad. Meanwhile Ford rehired 350 experienced engineers and topped the JD Power quality study for the first time since 2010. Vibes and "go faster" won't cut it any more, because we now have both the public failures and examples of what actually created leverage, and it was a system, not a catchphrase.

Two: a move to defining what is what. What is AI, what is human, and for anything in between, clarity about who actually owns it. Not "the agent did it" as an excuse, and "meat proxy" (someone who forwards AI output without reading, understanding or checking it) defined and ruled unacceptable.

The problem, and this is the reckoning, is that responsibility has been undermined across the industry. Treat humans like replaceable automations and that is how they will behave. The meat proxy is a systemic failure as much as a personal one.

My weekend release felt like a personal failure at the time, but in the week since I've tried to turn what I learned into a systematic solution. Another overhaul of the process, because we're shipping bigger features now that we've hit our groove on the new platform. I'd sooner break things and adjust than over-engineer the process up front.

What I've been doing

I'm continuing to write about building the software factory: how I measure whether ours is working, the full story of the Twill replatform, and the role of engineers in it.

Speaking of software factories, the Twill one is cranking up into a new gear as we finally start building the things the new platform has unlocked.

I ran a little experiment to try and reverse engineer code review, which came up in the Dev Interrupted roundtable. My spicy take? It's not really about code review. That's just always been where conflict on teams shows up.

We're wrapping up our first year of DRI Your Career, with cohorts in flight for DRI Your Career, EM Survival Guide and Breaking Into the Security Mindset, and working on exciting things for 2027. If you have dev budget left or know someone who does, early bird pricing for the January DRI cohort runs until the 4th of December, and super early bird for the February EMSG cohort until the 8th.

My public speaking is DONE for the year. I rounded things off with the Anthropic webinar with Jean and next.app devCon in Berlin. I was on the program committee for swiftCon there, and seeing abstracts I’d liked become real talks, including Ben Scheirman's talk on the billion row challenge in Swift. Super nerdy optimisations all the way down: mmap and multithreading eventually, but my favourite part was swapping out Swift's standard types for raw bytes and unsafe pointers. If this is your jam, he did a whole NSScreencast series on it.

Personally, my naturalisation certificate arrived (yay for a new passport, and getting back into the good queue), and I had a gorgeous and much-needed long weekend in Killarney with friends.

What I've been reading

• Sales Pitch by April Dunford - Like Obviously Awesome, just very clarifying and helpful.

• Walking on Broken Glass by Shannon Nutter and Julia Carreon - I really want there to be a book about the experience of women in leadership, but this isn't the one I was hoping for. It tried to be everything, so it ended up being quite shallow: a hundred interviews flattened into broad buckets, and advice that boils down to "document, lawyer up, get out." It had its moments, though.

  "The moment a woman deviates from the script - by asserting herself, setting a boundary, or calling out unfairness - she may be smeared as 'not a cultural fit', 'not a team player', 'self-centered', or worse. Fear is a powerful silencer. It's no secret that women who refuse to put up with nonsense often face social and professional backlash." - How tech workers are feeling in 2026 from Lenny's Newsletter - Burnout is up, optimism is down, and the biggest worry about AI isn't losing your job but being expected to do more for the same pay. My take: this is a result of pressure for AI to deliver results without any space to experiment and learn. - Relocating Rigor by Chad Fowler - I'm obsessed with rigour as the mechanism for real AI-driven velocity, so I enjoyed Chad making the case for it from a different direction: every shift that looked reckless just moved the discipline somewhere harder to fake. Also one of the more useful descriptions I've read of what the SWE job is turning into.

  “The engineers who thrive in this environment will be the ones who relocate discipline rather than abandon it. They’ll treat generation as a capability that demands more precision in specification, not less. They’ll build evaluation systems that are harder to fool than the ones they replaced. They’ll refuse the temptation to mistake velocity for progress.” - The Manager's Path in the Age of AI by Camille Fournier - "A lot of what we call 'management' is in fact a risk management function." This line from Camille's talk was a lightbulb moment for me. Obviously that's what managers do; I'd just never seen it put so simply, or noticed how much it covers. It's also her case for why technical skills still matter in management. It's much easier to manage those risks if you understand how the system works. - Write it yourself by Jon Seager - on going back to writing it himself. There's a difference between what's for machines and what's for humans, and we shouldn't juice our own productivity by wasting other people's time. I use AI on most of what I write, usually: rough notes in my own voice, iterate, edit, repeat. My blog posts are taking more time, not less, because the feedback loop is fast enough that I go round it more often instead of rereading the same draft and losing sight of how it can be better. Is it making my writing better? Sometimes. Sometimes it's helping me get something into words sooner. But one thing I'm fanatical about is information density. - Your values aren't a democracy by Kieran Snyder - If your values can't help you choose between two things you care about, they aren't really values. And if you want to know what a team actually values, follow the money. Look at who gets hired, who gets promoted, and who gets the raise. - How I built a to-do system that actually gets things done by Hilary Gridley and Trading Notes on Personal OS by Ellen Chisa - I'm fascinated by other people's productivity systems because of what they reveal about the inner workings of someone's brain, and both of these taught me something about mine. Reading Hilary's made me realise one way I use mine is to get different views of what I need to do depending on headspace. Ellen's personal OS feels designed; mine feels... organic. It grew one annoyance at a time, so it has the architecture of a hedgerow. Both of them get at the fundamental question of productivity systems: not "how do I organise it all," but how do I actually get things done?

Know someone who'd enjoy this? Forward it on - or they can subscribe at whatsmyjob.club 🦝