Role prompt

How a role prompt works

A role prompt puts in one text what a team tells a new person when they start working together: what role they play, where they start, how they report the result and what they leave to others. A coding agent with another prompt plays another role: one time it delivers tasks, another time it breaks material down into tasks. The team’s process is then written down, not repeated in each conversation.

Anthropic’s documentation advises setting the role in the system prompt, the instructions a Claude model gets separately from the conversation’s messages. According to it, a role focuses the model’s behavior and tone on your use case, and even a single sentence makes a difference.

An agent in a console can also get a role prompt while it works. The MCP protocol provides server prompts for this: the client shows them, for example, as commands, a person picks a command, and the server sends back the prompt’s content as a message that goes into the conversation.

Role prompts in featureloom

In featureloom prompts belong to the organization, not to a project. A prompt is one of your team’s rules or procedures in markdown, and every person, console and automation in the organization can read it. Every new organization gets two prompts: Realizer says how an agent delivers one plan from the board, from taking it to reporting it, and Planner how it turns material into plans on the board. You change them on the Prompts screen and write your own with the New prompt button.

A console fetches a prompt with one command: the word featureloom and the prompt’s command name, such as featureloom realizer. In Claude Code prompts work as commands. With a plan key, such as featureloom realizer INV-14, the console gets in one call the role prompt, the whole plan document and the documents of the delivered plans it directly depends on. The prompt reaches it whole and word for word. The agent can also look through the organization’s prompts itself and fetch the one it needs, because every console can read prompts.

An administrator can assign a prompt to a whole plan type on the Organization settings screen, in the Plan types tab. You assign a prompt to a single plan in the Console’s prompt field on the plan’s screen, and the plan’s own prompt takes precedence over its type’s. That way consoles can deliver plans of one type in different roles. A plan with a prompt waits for a console, and its card on the board says, for example, For a console: Realizer.

Who changes prompts follows from the rights. In the app every person creates and changes prompts, while changing the command name and deleting a prompt are left to an administrator. A console creates prompts and saves new versions of them when its person has the Edits prompts right, and it describes each change of content in a version note. Each save that changes the text makes a new version, and you can restore an older one from the history. featureloom does not run prompts itself: it keeps them and hands them to the console that asks.