Hey folks, I published a small agent skill called joedevflow.
It is an agentic software engineering workflow centered on TDD, separation of concerns, and deliberate context management.
The premise is that coding agents are more useful when the workflow keeps them constrained: one mode at a time, deliberate context management, and tests written before implementation so the agent has an executable target.
joedevflow separates agentic software work into four explicit modes:
- Design
- Test / Red
- Implement / Green
- Observe, Debug & Maintain
Each mode has a specific job. Design decides what should be built. Test mode turns that into failing assertions, proving the behavior does not exist yet. Implementation makes those assertions pass without editing the tests. Debug/maintenance verifies behavior, coverage, and regressions.
The skill also uses a HANDOFF.md file between modes, so each phase can leave behind a compact summary of what happened, what changed, what remains risky, and what should happen next. That makes it easier to move across context windows or even hand the next phase to a separate agent.
Install:
npx skills add JoeCardoso13/joedevflowRepo: https://github.com/JoeCardoso13/joedevflow
I’d especially like critique from people using coding agents for real implementation work. I’m interested in whether stricter mode separation, test-first implementation, and explicit handoffs improve reliability in practice.