Why use a workflow

A well-written prompt handles most of what you ask a coding agent to do. On a large feature or a multi-step task, though, a single prompt asks the agent to plan and carry out everything at once. That’s when it tends to drift: it loses track of the goal, or makes a sweeping change across the codebase that’s hard to review.

A workflow breaks the task into smaller steps and keeps the agent working on one at a time. You stay in the loop at each step, so you can check the agent’s direction before it commits to the next one.

You get these workflows as skills from our skills repo, so there’s nothing to copy and paste.

Note

Skills are a standard format for Markdown files that a coding agent can invoke — automatically when it spots a relevant task, or manually as a slash command.

For example, run the roast-me skill with /roast-me I want to refactor the licensing code to make it more reliable.

You won’t need a skill for every task, but they earn their place in the situations we cover over the next few pages.

I’ve adapted these skills from Matt Pocock, who has some great content on YouTube about software engineering with agents.