AI development workflow
AI is one of my tools, not a replacement for judgement. This is the setup I use across my projects, including this site, and the guardrails that keep it honest.
- Type
- Workflow
- Status
- In use
- When
- 2026
- Role
- Design and engineering
- Claude Code
- MCP
- Unity
- Notion
- GitHub Actions
Decisions worth a look
- Specialised agents with their own context: a level designer, a balance tuner and a feedback specialist, coordinated with explicit handoffs.
- Project rules live in the repository (AGENTS.md, prompt files and the .claude directory), so every session starts with the same conventions.
- MCP connections let the agent read and edit the Unity scene and the Notion design wiki directly instead of working from pasted context.
- Nothing merges because an AI wrote it. Type checks, tests, accessibility checks and performance budgets run in CI on every change, and I review every diff.
The loop
- Write down the intent in the design wiki or an issue.
- Let the right agent draft the change with the project's rules loaded.
- Run the tests locally and read the diff.
- Open a pull request; CI decides whether it can merge.