How we build
Seven stages, the same for every product.
A product does not move forward on enthusiasm. Each stage has a job, a check that must pass, and a record of what was done.
The method
From a problem to a product that keeps improving.
Discover
Understand the problem.
Research
Users, market, requirements.
Design
Experience and architecture.
Engineer
Build the software.
Verify
Function, reliability, security.
Launch
Release what passed.
Improve
Monitor, learn, refine.
At every stage, AI prepares and checks; a person decides. The table below says who does what.
Discover
We look for a job a business repeats, where the tools it already has leave a gap. A problem is written down in the words people use for it, with where it was found.
AI does
Reads public sources and gathers statements of the problem.
A person decides
Whether the problem is worth more work, and which areas we stay out of.
Record
Each claim, with its source and whether it was checked.
Research
Who has the problem, what they use today, what they already pay for, and how a product would reach them. If the evidence is thin, the idea waits.
AI does
Compares existing products and summarises what buyers say about them.
A person decides
Which product to build. This is never automatic.
Record
The buyer, the competitors compared, and the one way to do better.
Design
Requirements are written first, each one testable. Then the data model, the pages and the rules: what the product will do, and what it will refuse to do.
AI does
Drafts requirements and architecture from the definition of the product.
A person decides
Scope, pricing, and what the product must say about its own limits.
Record
Requirements, schema and the list of every route.
Engineer
The product is built on a common, tested foundation. Only the logic that is the product's own job is new code, and it is written together with its checks.
AI does
Writes and revises code and tests within its role.
A person decides
Design questions the requirements did not settle.
Record
A numbered build: the exact files that will be checked.
Verify
Automated checks, a security review, and acceptance against the written requirements. Then a scripted visitor uses the product end to end in a real browser.
AI does
Runs the checks, reviews the code, and reports findings.
A person decides
Nothing here: a failed check stops the build. No one can wave it through.
Record
A report for each stage, with every check and its result.
Launch
A release is a fixed set of files with a manifest. It goes live only when a person approves it, and it is tried as a customer would before the address is announced.
AI does
Prepares the release, the runbook and the rollback steps.
A person decides
To publish, to spend, and what the product may claim.
Record
The manifest of the release and who approved it.
Improve
We watch whether the product is used and where people stop, read every support request, and change the product on that evidence. Each change goes through the same stages.
AI does
Summarises usage counts and drafts the changes that follow.
A person decides
What to change, and anything about a customer's money or account.
Record
Dated changes to the product.
In practice
What this looked like for our first two products.
- Commission Desk
- The release now live passed 151 automated checks and a 44-step browser run. It opened on a free plan on 17 November 2026.
- Reminder Desk
- Before opening, it was connected to a QuickBooks test company. It sent one reminder, and withheld two when the invoices were paid moments before sending.