I've been spending nights wrestling with Cursor/Cline/Windsurf on anything larger than a toy repo:

- the LLM cheerfully deletes my hand-written files - integration bugs that surface only after three merged PRs - no shared mental model between me and the agent

So I built Blueprint, a tiny desktop tool that lets humans and AIs agree on a plan before anyone touches code.

How it works

- Shared plan.md – human & LLM edit one file. - Auto-generated graph.yaml – turns the plan into boxes & arrows. - Per-component agents – each LLM only sees its slice; lock any file you want protected.

100 % local (Svelte + Tauri + Rust) and MIT-licensed. Alpha quality—expect sharp edges—but the core loop (plan → graph → generate) is functional.

I dog-food it on a toy collaborative editor I’m writing in C. First time I handed the repo back to the LLM without Blueprint it trashed my hand-rolled CRDT. With Blueprint it only touched the component I asked it to, left my locked files alone, and the diff actually compiled.

Would love brutal feedback or war stories from anyone else trying to get LLMs to play nicely on real projects.