How I work
I use coding agents in every step of the development cycle, inside a process that keeps the work specified, bounded, tested and reviewed. Judgment and accountability for what ships stay with me.
The process
01
Spec first
Every change starts as a written spec: the problem, the scope, what stays out and how to verify the result. Agents work from the spec instead of a chat history.
02
Plan and bounded tasks
The spec becomes a plan and small tasks, and each task is checked when it is done, so an agent only holds the context one task needs.
03
Implementation by agents
Coding agents implement one task at a time, inside the repository's instructions (AGENTS.md): conventions, commands and the rules they must keep.
04
The same gates for everyone
Every change passes the same gates, whoever wrote it: typecheck, lint, unit and end-to-end tests, accessibility checks and checks on the built output.
05
Human review
I review every change and merge only what I can explain. Accepted findings go back to the agent to fix, and the change is reviewed again.
06
Decisions in ADRs
Decisions that shape the system are written down as architecture decision records, so the reasoning outlives the session that produced it.
Where to see it
This site
Its public repository shows the agent instructions, the architecture decision records, the tests and the CI gates every change passes.
sdd-agentic-flow
The process as an open-source harness of skills for Codex, Cursor, Claude Code and GitHub Copilot.
AI in products
The same rule for AI features: AI speeds up the flow, guardrails and people make the decisions.