Sometimes, /implement-spec performs poorly, but using /implement with sub-agents yields better results.

[leogolfeto]

Hi folks! I've been using /implement-spec for the AFK part of my development workflow since version 1.3 became available. Before that, I had created my own skill to solve the same problem. Here's its prompt:

Act as an orchestrator of agents. Each sub-agent must use the `implement` skill to complete a task from the spec. Sub-agents should also spawn their own sub-agents for tasks such as code reviews.                               
                                                                                                                                                                                                                                    
 Your role is to manage task dependencies and ensure that each sub-agent runs the `implement` skill. Once a task has been implemented and reviewed, commit the code and mark the corresponding task as completed in the issue tracker.                                                                                                                                                                                                                          
                                                                                                                                                                                                                                    
Use `gpt-6-luna` with medium reasoning effort for each sub-agent. Each sub-agent must start with a blank context and invoke the `implement` skill with the specific issue assigned to it.              

In my POV, my custom skill performs better than /implement-spec. I think the main reason is that it invokes /code-review through a nested sub-agent after every task, while /implement-spec runs it only once, after all tasks have been implemented.

This might be particularly helpful with my setup: I use a stronger model, such as gpt-6.1-sol or another model in the same class, for orchestration, and a smaller model, such as gpt-6-luna, to implement each task.

Have you considered running code reviews after each task instead of only once at the end?