Blueprint: A planning copilot that one-shots bigger coding tasks

We spent the last dev cycle on a planning problem: coding agents are fine on small, defined tasks. On anything ambiguous, they rush to code, invent assumptions, and often leave a lot of cleanup work.

Claude Code's plan mode asks questions to unblock itself, which is helpful. Spec-kit and similar tools generate a plan and ask you to review it, which is OK. Neither puts the human decision upfront.

So we built Blueprint: https://github.com/imbue-ai/blueprint.

Before any code is written, it reads your repo and runs rounds of multiple-choice questions about the task. You answer what matters and the agent surfaces things you hadn't considered. The output is a plan you can hand to Claude Code, Codex, Gemini, or Pi.

Blueprint works well across a variety of tasks and codebases. Our beta testers used it for greenfield projects, complex features in mature codebases, research experiments, and more. It's less suited to frontend development, small refactors and debugging.

Blueprint is agent-agnostic, free and open source. It's available as agent skills or extensions in VS Code, Cursor, and Windsurf.

Happy to talk about the architecture, the question-generation prompts, or anything else. Eager to hear what the community here thinks!