Roo Code, But Your System Prompts Are in YAML

What it does

RooFlow provides five specialized “Flow” modes—Flow-Architect, Flow-Code, Flow-Debug, Flow-Ask, and Flow-Orchestrator—that sit alongside the standard Roo Code AI assistant personas inside VS Code. It layers on a persistent Memory Bank, adapted from an earlier memory-bank project, so the assistant retains project knowledge across sessions. The prompts are written in YAML rather than Markdown, with the stated goal of improving efficiency and reducing token consumption. The README is upfront that this is not officially endorsed by the Roo Code project.

The interesting bit

Flow-Orchestrator acts as a dispatcher, delegating complex tasks to the other four modes while simpler tasks can target them directly. A Python installer can also scrape your existing Roo Code system prompt—including connected MCP server details—and inject it into RooFlow’s YAML templates, though re-running that installer wipes any manual edits you made to the generated .roo/ files.

Key highlights

• Five integrated modes with distinct responsibilities; Orchestrator delegates to Architect, Code, Debug, and Ask
• Memory Bank access controls: read/write for Architect, Code, and Debug; read-only for Ask
• YAML-based system prompts stored as .roo/system-prompt-[mode] files
• Optional MCP server tool discovery by parsing a copied Roo Code system prompt into YAML
• Installer scripts handle cloning and templating, but overwrite local customizations on re-run

Caveats

• Explicitly experimental and not officially supported by the Roo Code project
• Re-running the installer overwrites the .roo/ directory and .roomodes, destroying manual edits
• Claims of reduced token usage and improved efficiency are aspirational; no benchmarks are provided

Verdict

If you are already living in Roo Code and want to tinker with prompt formats and session memory, this is a curiosity worth installing. Everyone else can safely wait for these ideas to surface in the official extension.