01Method
How I build
A duck swims, runs and flies, and is the best at none of the three. In Brazil that is what you call a person like me.
It was the joke, and it was fair: technology, product, development, design, business, a lot of ground covered and never the specialist anyone called for. The joke stopped working when breadth turned into the qualification. It is what lets me take a problem from a business question to a running system without waiting for five other people to be free.
Writing the implementation stopped being the bottleneck. What replaced it is harder to fake: knowing the domain, knowing how to research it, planning before building, and writing it down in a structure a model can be held to.
02Toy or work tool
AI without organised, structured, standardised data is nothing. Without governance, process, conventions and a team trained to use it, AI is a toy in the hands of whoever happens to feel creative, not a work tool. That is the job: the ground under the model, not the model.
I have watched a small team deliver in weeks what would have taken months: APIs, applications, MCP servers, command-line tools. The constraint stopped being how many specialists were available to hire.
And the gain has a ceiling. I simulated the same application built by teams of increasing size, and the curve turns: past a point, adding people stops accelerating the work and starts delaying it. Producing code was never the bottleneck. Absorbing it is, because reviewing, integrating and testing stay bound to the number of people, and the team size that works is the number of independent domains in the system.
03Four disciplines
- Master the domain
- The domain goes into the repository as readable doctrine, not into the head of whoever is prompting. A rule a model can read is a rule a model can be held to.
- Research
- A decision is recorded together with the alternative it beat. Preference is not a reason, and six months later nobody remembers which one it was.
- Plan
- A design spec paired with an execution plan, written before the code. The plan is where disagreement is still cheap.
- Document and verify
- Imperative rules, schema contracts, hooks and gates that fail the build. A harness does not ask the model to behave. It makes deviation detectable.
04The metamodel
I wrote the DDC Framework to stop rebuilding that structure by hand on every project. It is grounded in Domain-Driven Design and in the patterns I was already using, and it separates what an agent knows from how an agent acts: four layers of doctrine (business, product, engineering, operations) read before anything is written. It covers data and machine learning as well as product code, because a pipeline needs conventions at least as badly as an interface does.
05Measured, not asserted
That is a claim, so here is what it stands on. Counts taken on August 25, 2026 across 26 repositories.
26
Repositories in one year
24
Carrying a harness
14
Carrying a doctrine tree
236
Decision records
770
Design specs and plans
74
Enforcement hooks
06The number I am not printing
The doctrine travels. Across these repositories there are 279 rule files, but they are 66 distinct rules, replicated project after project. Publishing the larger number would have been arithmetically defensible and a lie. Replication is the interesting part: doctrine that moves into a new repository unchanged is an asset rather than a habit.
| Artifact | Files | Distinct |
|---|---|---|
| Rules | 279 | 66 |
| Skills | 694 | 203 |
| Subagents | 216 | 107 |
| Context documents | 1,244 | 582 |
07Two that make the point better than the totals do
One of these repositories holds no product code at all: 29 decision records, sequence diagrams, and the portal that renders them, an architecture documented before a line of the system existed.
Another ships an executable test suite written by three agents in sequence (planner, architect, implementer) from a description of the feature. Neither of them is explained by knowing how to code.
08What was counted
The figures above are counts, taken on disk, repository by repository. 13 of the 26 carry git history anything can read, and that is the set any commit figure of mine is summed over.
A model is worth the ground it stands on. Building that ground is the work.