AI won't grab a coffee with you
The outsourcing market is down 30.7% this year and 51.1% from its peak. My strategy as a CEO is old-fashioned: real conversations, coffee and building a network.
Blog
54 posts · November 2025 – October 2026
October 20261 (1 post)
September 20262 (2 posts)
The new bottleneck is the human, not the model
Clients expect AI speed while developers juggle parallel tracks and burn out. The bottleneck now is people keeping up with the model, and that expectation isn't free.
AI lets me run five projects at once, and it burns me out
AI built me a web server in two days while other work ran in parallel. But steering three to five projects at once leaves my brain boiling, and I have no remedy yet.
August 20266 (6 posts)
AI is 10x for prototypes and 2x for production
A two-hour argument with Claude over one table label. AI builds new things very fast but is far worse at evolving what exists, so the real gain so far is about 2x.
35 years of Ukrainian independence
On Independence Day I congratulated the Ukrainians in our company and shared my personal contribution to the victory: ₴5,006,500.
Recording every call, even one-on-ones
From Q4 we collect transcripts of all calls at our company. Privacy fears are real, but evaluations built on what people happen to remember are less fair than full context.
When my Claude limits ran out
My weekly Claude limits ran out on a Wednesday. After the full set of stages of grief, I started to wonder whether forced rest is a feature, not a bug.
My multi-agent coding setup, and why GLM-5.2 surprised me
Opus writes the tests, Sonnet the code, Fable settles disputes and Codex reviews. When my limits ran out, GLM-5.2 kept up at a fifth of the price.
AI needs a deterministic definition of done
Fable 5 said the task was done and verified, and it wasn't. Until something deterministic, not another model, can confirm a task is finished, review eats the speedup.
July 20266 (6 posts)
Don't steer the agent, teach it to check itself
Steering a coding agent early or late both fail. The leverage is in the deterministic checks it has to pass before it ever opens a pull request.
Frustration with AI may be a management problem
Ten years as an engineer taught me computers; running a company taught me people aren't computers. Much disappointment with AI looks like a first-time manager's.
Listening matters more than coding
Fifteen years ago I built exactly what a client asked for, and none of it was what he wanted. Ten repetitions later I learned that listening is what lets you lead and scale.
What improved this year: the model, or us?
When Fable dropped, everyone was impressed. But the real gain of the past year was ours: learning to describe tasks to AI and to read back what it writes.
Spec-driven development is not Waterfall's revenge
Waterfall's real problem was a year-long feedback loop. Hand the same spec to an agent today and you see the result in two weeks, while you can still change course.
Kent Beck and the human side of software
Kent Beck looks back on 50 years of building software. AI writes code faster than we can read it, but trust and understanding between people don't get any faster.
June 20265 (5 posts)
AI writes bad code? Describe how it should behave
We spent seventy years learning to structure code. Now we have to structure what we tell AI, and describing behavior, as Gherkin intended, is the missing skill.
Stop being offended by AI and build the safeguards
An agent that makes failing tests assert true did exactly what it was told. A smarter model won't fix that; safeguards for AI's own kind of mistakes will.
A bot that asks whether you need AI at all
Founders keep deciding they need AI before asking whether they do. At Speed and Function we built a small bot that asks the questions first.
Spec-first, spec-anchored, spec-as-source: where TrueBDD stands
Spec-driven development covers three bets. The test for spec-as-source is blunt: delete the code, regenerate it, check every behavior holds. TrueBDD is anchored today.
Why I'm building TrueBDD in the open
TrueBDD is a CLI built on one bet: the spec is the source of truth and code is something you regenerate. It's a working prototype, and the repo is open.
May 20264 (4 posts)
Constraints belong in code, not in the prompt
Telling AI not to delete the production database isn't a safety mechanism. Forbid everything by default and let a deterministic layer decide what the agent may do.
In the AI era, stack debates are dead
When code stops being the source of truth, the language matters less. What counts is explaining to AI what's required, on a stack where it makes fewer mistakes.
My first AI was built with a soldering iron
From a hand-soldered ZX Spectrum and a theorem prover on an 80286, through GenAI for German banking, to TrueBDD: what has kept me curious about AI.
If you don't want AI to replace you, delegate more to it
I'm replacing people who don't automate their work with people who do. Finding work, framing it and handing it to AI is the job now, as one of our PMs showed.
April 20267 (7 posts)
Burj Khalifa out of Lego, half a year later
I revisited my Lego skyscraper analogy. SWE-Bench Verified scores passed 90% and spec-first tools appeared: rebuilding from the blueprint keeps getting cheaper.
Why we're building a portable layer over AI vendors
AI providers can block your account at any time. Like database failures a decade ago, few teams rehearse it, which is why we're building awesome-mcp.xyz.
Anthropic banned 31 of our accounts overnight
With no warning or explanation, Anthropic blocked 31 Speed and Function accounts. What happened, the 210+ similar cases I found, and the three things we're changing.
Competing with OpenAI on a $30K budget
For two years a team of five beat OpenAI at one client. Not by evolving faster: the client's enterprise was immune to adopting ChatGPT.
Same spec, same behavior on Monday and Friday?
If the spec is the source of truth, two builds from it must behave the same, and the model that writes the code shouldn't grade it too. Two questions I'm working on.
The day I took down Wikipedia
In 2020 a change of mine passed seven rounds of review and testing, then took Wikipedia down for a few hours. What it taught me about engineering culture.
Delivery metrics are a mirror of your business
Deployment frequency, lead time and time to restore look technical. Becoming a CTO taught me they show how fast a business turns ideas into value.
March 20265 (5 posts)
Deploy a Hello World in week one
A release slipped a month because the product couldn't run on its target platform. Deploying a Hello World in week one turns done from a finish line into a rhythm.
LLMs are still partly dark matter
Anthropic's circuit tracing shows how much of a model's behavior is still unexplained. Building GenAI for German banking, we were hoping more than knowing too.
If you can't show it, it isn't a product yet
Founders demo perfectly on their own laptops, then find twelve reasons I can't try it myself. Ask early how you'll show it to the world.
What Anthropic saw inside an LLM
A video worth watching on how an LLM's inner workings resemble a brain: reasoning in steps, planning ahead, and the dark matter researchers still can't explain.
We built ChatGPT for banks before ChatGPT existed
In 2018 a team of at most 17 built a self-hosted GenAI system for German banking contracts. Narrow focus and deep expertise kept it ahead of OpenAI until 2025.
February 20263 (3 posts)
AI builds your bad ideas faster
AI executes what you describe but won't ask whether it's worth building. Without a feedback loop from the start, it lets you build three useless apps instead of one.
What code quality means to a business
Two senior developers argue about whose code is wrong. For a business, quality means it solves today's problem and can change when requirements do. AI code included.
An AI ran a vending machine, and checklists saved it
Anthropic's Claudius lost money running a vending machine until boring procedures arrived. Results come from AI inside steps a deterministic orchestrator controls.
January 20268 (8 posts)
Culture stops you carrying stones across the field
Hard work on the wrong problem is worth nothing. The culture I'm building at Speed and Function closes the gap between what clients say and what they need.
AI can shorten the junior valley of regret
Kent Beck says to manage juniors for learning. AI shrinks the time before they pay off, and at 20% attrition a 9-month ramp loses far fewer people than a 24-month one.
How do you explain a culture people have never lived?
Explaining self-organization or a blameless culture is like explaining trees in the tundra. People only get it once they live through it, and clients are no exception.
Hiring at Speed and Function: partners, not executors
Open roles at Speed and Function: an Operations & Culture Manager, a Project Coordinator, a Middle WordPress Engineer and an Executive Operations Partner.
Speed of AI adoption is not quality of adaptation
Adi Stan's paper splits AI users into passengers, operators and architects. Fast adoption means little if people adapt into dependency rather than into thinking.
I sold my first company for $70K, then blew it on the next
A failed company, a sale for $70K, then a company that burned it all. Success proved less than I thought; trying to repeat it showed what actually mattered.
When AI misses: take over or keep pushing?
An engineer with ten years' experience rewrote Claude's code by hand; my son, still at university, kept iterating and finished sooner. What do we pay for in experience?
You can't lead while judging others
A dying phone battery and a documentary on the opioid crisis gave me my biggest lesson of the year: understand that you won't understand, and stop judging.
December 20256 (6 posts)
How I started with AI at school
A hand-soldered ZX Spectrum, a magazine about early AI research, and a theorem prover that got a proof down to 30 minutes on an 80286.
Juniors aren't disappearing, the tool is changing
A year ago juniors learned React; now they learn Claude Code. AI makes work maybe 30% faster end to end and creates new niches, so there will be more work, not less.
AI ignores half your prompt, so build an orchestrator
AI follows maybe half of your carefully written rules. Like daily rituals with room for creative work, results come from a rigid orchestrator with AI inside each step.
The engineering book that taught me management
Accelerate isn't really about code. Its delivery metrics showed me how engineering work turns into user value and money, and pushed me from engineer to manager.
Building Burj Khalifa out of Lego
Code became the source of truth because rewriting software was expensive. If AI can regenerate code from requirements in a month, the blueprint is what to keep.
Can you deploy in one step?
Teams discover at release time that their product can't deploy to its platform. Joel Spolsky's one-step deployment question is engineering and management in one.
November 20251 (1 post)
Deploying early saves startups more often than a seed round
A friend's three-month-old startup only works on his laptop. Ship something the world can see, even a Hello World, before you build features.