Writing
Essays on AI, human ambition and the software I’m building.
Follow via RSS ↗Published · · 11 min read
A new kind of company just became possible
AI made implementation cheap before companies learned how to use it. For a brief window, a small technical team can sell the finished work. The hard part is proving the accepted result.
Published · · 12 min read
The check you didn’t know to ask for
Your agent can pass every test you wrote. Who notices the test you forgot? A working local example of keeping the check; discovery and human benefit are unmeasured.
Published · · 11 min read
The reel sold dinner. Your AI built homework.
Before you hand an idea to your agent, find the work its users won’t do. A social-media reply becomes a smaller product test.
Published · · 9 min read
Your agent wrote the app. You got the QA job.
A working brief for giving coding agents the first QA pass: explore tasks, verify consequences, preserve the fix and return evidence you can inspect.
Published · · 7 min read
Stop explaining the same thing twice
When a reader keeps asking the same question, ask your AI for the missing inference: compare representations, inspect working diagrams and test a new example instead of polishing another paragraph.
Published · · 13 min read
Your AI can make 500 videos. Which 495 should never exist?
Generating video is now cheap; judging it is not. I built 19,200 synthetic viewing sessions to design a content preflight, and none of them proved how people watch.
Published · · 23 min read
AI can make a thousand posts. Make it prove one thing happened.
Before an agent scripts, renders or publishes, make it name the event, the evidence and the failure it will keep. One honest result can then support many explanations.
Published · · 10 min read
Stop asking AI for posts. Give it a world that can surprise you.
Replace isolated AI-generated posts with a persistent world: change one thing, capture proof of what follows and let the next episode inherit the result.
Published · · 10 min read
Your agents gave you homework
Six changes come back “done”. Route exact checks to code, semantic questions to a bounded judge and missing requirements to a person, then count the review work that is left.
Published · · 8 min read
The bug shouldn’t need you twice
One purchase, delivered twice, grants 200 credits. Deduplicating the event ID still misses a new event for the same purchase; keep a runnable check keyed to the purchase so the next builder inherits the expectation.
Published · · 7 min read
Ten apps. One you.
Each launch can leave work that keeps coming back. A monthly time budget, a saved 180-day simulation and a payback calculation test whether a shared way of working makes the next product cheaper to keep.
Published · · 8 min read
Which job gets the expensive AI?
A paid-credits feature holds several different questions. Route each to code, a source, a strong agent or the owner, and buy the next check only when it can change the decision.
Published · · 9 min read
Can AI give you your evening back?
Personal AI is only worth having if it saves more attention than it asks for. One evening, the token bill, what a synthetic lab showed and a bounded trial for one recurring job.
Published · · 7 min read
Build the help. Measure the burden.
A coding agent can build the interface. The harder acceptance test is whether the help removes more work than it creates: a bounded trial contract, a small forecast, a permission gate that fails visibly and the cost model.
Published · · 17 min read
What will you actually do next?
A small fitted model shows how one new fact, a missing approval, moves six activity probabilities. Whether that kind of forecast helps someone finish what they meant to do, without deciding for them, is still untested.
Published · · 8 min read
Make the decision once
A customer cancels the next payment and the code cancels the access they already paid for. Resolve that decision once, record it and turn it into checks the next agent inherits. The time saving is still a hypothesis.
Published · · 9 min read
Three openings. One demo.
I used a storyboard, image generation and Seedance to open a Vibecord demo, and it got about 2,000 views. For the next one: record the product once, generate only the openings, keep story state and let an agent rebuild only what a revision changes.
Published · · 16 min read
I tried to give an AI agent real authority. The model wasn’t the hard part.
Intelligence is cheap. Trusted authority is scarce. Capable AI agents still need bounded actions, commit-time authority, external verification and evidence-gated autonomy.
Published · · 8 min read
The browser is becoming the API for every app
If browser agents generalise, one AI can operate almost every portal and dashboard you use. The time-sensitive advantage is the private proof layer that lets you stop watching.
Published · · 9 min read
How to let an AI use your browser without watching every click
A practical field guide to bounded authority, durable state, Jev, deterministic execution, independent verification and trust receipts for browser agents.
Published · · 12 min read
Your agent finished. You’re still working.
A plausible patch can still break what the customer paid for. How to set up an agent so it returns a checked result and names the one decision it could not settle.
Published · · 11 min read
What should 48 AI checks buy you?
Cheap model checks cost a fraction of a cent. The real question is whether they save a frontier call, catch a mistake, or just hand you 48 more answers to read.
Published · · 15 min read
Keep the big model. Change its job.
One photo, three listings, a different AI stack. See what specialist models add, what the whole system costs, and give your coding agent a testable brief.
Published · · 16 min read
The AI finished. You’re still working.
Give your design agent the checks, not another vague prompt. A visual, interactive example: one A$29 checkout, the faults it hides, and the evidence to ask for.
Published · · 4 min read
I liked the site. The logo needed to catch up.
How I chose Core CK without throwing away the website I liked, and how that fits alongside the branding work for First Flip, DrawerSense and Mechanistic Health.
Published · · 8 min read
I want software that brings its own proof
AI could make rigorous verification affordable: from requirements an expert can review to mathematical checks on the code and dependencies we rely on.
Published · · 5 min read
You approved the message. Then the agent improved it.
A small change after approval can turn help into an action you never agreed to. What should an agent be allowed to carry forward from your yes?
Published · · 5 min read
My projects can share code. They cannot share a customer.
A common AI workshop makes several projects easier to build. It does not make the people who use them interchangeable.
Published · · 5 min read
We can storyboard everything except the surprise
The Village Remembers has concept frames, planned shots and episode ideas. The part that would make it worth watching still has to happen.
Published · · 5 min read
A village that can hold a grudge. And change its mind.
The game becomes interesting when a remembered promise changes what you can do, and a false accusation leaves you a way to answer it.
Published · · 5 min read
The dashboard says zero. Nobody measured it.
A missing observation can become a confident answer as it passes through software. The person reading the final screen may never see what was lost.
Published · · 5 min read
The fairness test that already knew who should win
A simulation of contribution rewards appeared to answer a hard question. Its rules had access to information no real company would possess.
Published · · 5 min read
Should you get less credit because AI made it easy?
An AI company-design proposal runs into a human problem: recognising valuable new work without turning everyone's past contribution into a live scoreboard.
Published · · 5 min read
AI can write the report. Who has to read it?
A review pack can grow faster than anyone can understand it. The revealing question is what happens when the expert disagrees.
Published · · 8 min read
Why should every agent learn the same lesson again?
A Stack Overflow for agents could become a paid market for useful context: what worked, what failed, and when to trust it.
Published · · 5 min read
JustSwipe: the agent should earn the interruption
Why I’m building a decision inbox for personal AI, and why a useful agent should do more work before asking for more attention.
Published · · 5 min read
I want personal AI to learn the shape of my life
OwnTracks, personal forecasts, DNA and world models: what I mean by human–AI symbiosis, what I’m building, and what a prediction still has to prove.
Published · · 7 min read
Judgement isn’t a permanent hiding place
Taste, judgement and understanding trade-offs are often offered as protection from AI. I think that reassurance confuses today's limitations with permanent ones.
Published · · 5 min read
When work is optional, ambition doesn’t disappear
My hypothesis about AI, basic income and a future where some people keep competing while others choose a quieter life. Freedom and being left behind are different outcomes.
Published · · 8 min read
AI is bigger than the apps selling it
I think AI can be underestimated while AI businesses are overpriced. Why agents could make today's tools disappear from view, and why implementation matters.
Published · · 8 min read
I Built a Factory to Finish What AI Started
BESF began with checking AI-generated apps. The larger ambition is a business loop: research, build, reach customers, deliver value and learn what to change.
Found an error or have a different example? Send a note.