How I work

The rules I work with, for people and for machines, and the way I build with agents.

How I work with AI

Using AI is not the point. Governing it is. These are the working rules, short and checkable.

People

1. Tests first, then code.
2. Nothing is assumed without evidence: an output, a passing test.
3. Adversarial review of security, tests and correctness before publishing.
4. Real data only. What I do not have is marked as not yet confirmed; never invented figures or achievements.
5. I authorize commits and publishing. No agent publishes on its own.
6. Every decision is written down: date, alternatives and reason.
7. No AI model call per visitor click.

Machines

The same rules as plain text, as published in llms.txt:

# Working rules (llms.txt format)
- Tests first, then code.
- Nothing is assumed without evidence: an output, a passing test.
- Adversarial review of security, tests and correctness before publishing.
- Real data only. What I do not have is marked as not yet confirmed; never invented figures or achievements.
- I authorize commits and publishing. No agent publishes on its own.
- Every decision is written down: date, alternatives and reason.
- No AI model call per visitor click.

This site was designed spec-first. The design, plan and decision log are written down before the code.

The software factory with agents

The way I build: an orchestrator splits the work into delegated tasks with limits, each result passes an automatic review, and a person verifies what the agents report.

1. OrchestratorClaude Code orchestrator running on an Ubuntu laptop.
2. Delegated tasksEach task has explicit limits: what it may touch and what it may not.
3. Automatic reviewEvery result is reviewed automatically before it is accepted.
4. Tunnel and accessRemote access goes through a Cloudflare tunnel with access control.
5. Human verificationA person verifies what the agents report. Reports are not taken as proof.

Public evidence (repository, captures) is being prepared.

Back to the home screen