Spec-driven AI development
AI development,
on our own workflow.
From writing the spec to implementation and pull requests, the work is standardized with agent skills we built ourselves. Cross-review by a different model and steps where a person decides are built in.
Adversarial review: flag, fix, verify again
More on the development workflow
We generate the spec from business requirements, then implement against the agreed spec and acceptance conditions. Our skills define how specs are written, the coding rules and how pull requests are made. Different models, such as Claude Code and Codex, take the generating and reviewing roles and repeat flagging and fixing.
Hooks we built control the steps, and the points a person must review are set in the workflow. Work runs in parallel on Git worktrees, and changes verified by CI are merged.