Will you tell me where AI was used?
Yes. It's used everywhere, and I don't hide it. The diligence pack says what the agents wrote, who reviewed it and what got rewritten.
How I work
Agents do the typing.I make the calls, check the workand put my name on what ships.
This is the whole workflow, out in the open. There's no secret sauce. When you're ready, I teach your team to run it.
You get to see what the AI wrote, what I threw out, and why.
The agents do the typing.
- Write most of the code.
- Run the automatic checks.
- Dig through the research and draft the write-up.
- Take the small, well-defined jobs, several at a time.
An agent is an AI that can read your project, write code and run it. They're fast, they don't get bored, and they sound exactly as sure of themselves when they're wrong as when they're right.
I do the deciding.
- Decide what gets built, and what doesn't.
- Read every change they hand back.
- Throw out what's wrong.
- Use the real thing the way your customer would.
- Put my name on what ships.
That's the part you're paying for. The typing got cheap. Knowing when the machine is wrong didn't.
Five rules thatkeep the AI honest.
They're on my homepage as one-liners. Here is what each one means when it's your product on the line.
Types and unit tests passing is necessary. It is not sufficient.
Automatic checks are the floor. They tell me the code does what the code says. They don't tell me it does what your customer needs. So once the checks go green, I open the real thing and use it the way a customer would. One time every check was green, I did that anyway and found about fourteen more bugs.
For the technical reader: a green type check and unit suite gets a change to review. It doesn't get it shipped. I walk the real flows by hand before anything goes live.
Verify as you prompt. Prune as you prompt.
I check each piece as it comes back, not at the end when there's a mountain of it. Anything wrong, extra or unasked-for gets cut on the spot, before the next step gets built on top of it. An AI will happily slip in wording nobody approved, so I read every change.
The scratchpad is not the document.
An AI's working notes are full of guesses written down like facts. I've watched one invent a 72-hour grace period for late payments and state it like a rule. Nothing moves from its notes into something your team or your customers will read until I've checked it against the real source.
Five cheap narrow agents beat one long tired thread.
An AI that has been in one long conversation for hours starts forgetting things and cutting corners, a lot like a person at the end of a double shift. So I break the work into small jobs and hand each one to a fresh agent with exactly one thing to do. The work comes back cleaner and it costs less.
For the technical reader: short-lived agents with narrow context and one task each, instead of one long session that has to remember everything.
Nothing ships until a person who can defend it says so.
That person is me. Ask why something was built the way it was and you get an answer from the one who made the call. You never get a shrug and "the AI did it."
Where the agentsdon't get the last word.
I use AI on everything. That doesn't mean it gets to decide everything. On these, a person signs off every time, and that person is me.
Anything your customer reads.
Emails, prices, promises, the words on the screen. A made-up promise in a customer email is still your promise.
Anything that touches money.
Billing, refunds and payments get read line by line. An AI will invent a billing rule and write it down like a fact.
Who can see what.
Logins, permissions, passwords and keys. An AI will tell you the database is locked down. I go and look.
The tests that prove it works.
Sign-up, payments and permissions get tested end to end, and no agent gets to edit the answer key. Sometimes the tests pass because the AI rewrote the tests.
Whether to use AI at all.
If AI is the wrong tool for the job, I'll say so. Sometimes the fix is a cleaner process, not a smarter robot.
Every build ships with the foldera buyer asks for first.
I call it the diligence pack. If you ever raise money, sell the company or hire your next engineer, this is what they'll want to open. It's also how you check my work without taking my word for it.
- Ownership
- The code, the cloud account and the domain are in your name. You own the work outright.
- Architecture notes
- One page and a diagram that explain how it all fits together, plus an instructions file so the next team's AI agents can maintain it.
- Security check
- Logins, who can see which data, passwords and keys, and the third-party code it depends on. Read by a person, with the findings written down.
- Tests that matter
- End-to-end tests on sign-up, payments and permissions. None of them edited by an agent to make itself pass.
- Runbook
- How to deploy it, roll it back and change the keys, and who to call.
- AI disclosure
- What the agents wrote, who reviewed it, and what got rewritten.
- More than one driver
- At least two people who can deploy it, and a list of who has access to what.
- Billing path
- If it takes payments, the billing events are wired up and reconciled.
Straight answers.
What about my data?
We talk about what kind of data is in there before I look at any of it. For a first look, black out whatever you can. Real customer data only goes through tools we've both agreed on.
Why not just vibe-code it myself?
I do it. There's a time and a place for it. Production code that ships to your users or your customers and makes your first impression is not it.
Am I stuck with you forever?
No. When you're ready to take my fees off your overhead, I hand you the keys and teach your team the workflow that runs it. I stick around after if you want me to.
The short version of why I work this way is in the manifesto.
Want to watch it work?
Bring one real problem.
Book the free half hour
The first half hour is free. I'll tell you straight whether you need me.