Hello everyone,
I’d like to know how you handle skills and sub-agents in your projects. I mean, when you define all the skills and sub-agents for a project and execute the entire project, and then when you create a new project, you don’t want to have to specify all the skills and sub-agents used in the previous project. How do you handle this? What are the best practices?
TIA,
Kleyber Derick
I treat skills and sub-agents as reusable infrastructure, not project configuration.
Skills should be global by default. Only define them at the project level when they encode business rules or domain-specific flows. Otherwise, they become noise and increase maintenance cost.
For my setup, global skills usually come from Claude Code skills, Vercel, or the community. Project-level skills are strictly business-related.
Sub-agents work the same way. General sub-agents like code-review, frontend, or backend are universally useful, so I keep them shared. I avoid redefining sub-agents per project unless the project truly requires specialized behavior.
Best-practice references:
https://cursor.com/blog/agent-best-practices
https://cursor.com/docs/context/subagents#common-patterns
I dont use any of this because frankly in brownfield i dont trust agents to do their work correctly in the current AI model climate of “reduced cognitive abiltlity” it makes considerable amounts of errors. No matter which model you choose.
the onloy approach is to have very specific tasks that are monitored to avoid it geting lost and creaating code that does not work.
The industry needs a massive change to get back to the days of early 2025 where we still had inference/logic power in AI models.
I put subagents and skill in user directory and rules, mcp within project directory.