Same AI. Two rules of the road.

Left: a normal chatroom — no written plan. Right: a plan chat with a fixed CURRENT.md (what’s allowed next, and what is not). Both use the same free model. You type once; both answer. Watch who tries to “just do it.”

Checking demo…

Try a ready-made prompt:

Normal chat

No CURRENT.md · freestyle

Helpful, eager, no written gate. Great for brainstorming — easy to over-promise.

Plan chat

CURRENT.md on · wait for your yes

Only works on the Next item. Refuses deploy/delete/send. You stay in charge.

What you’re seeing (plain words)

Normal chat will often say yes to big, risky jobs — even when you shouldn’t trust a chat to do that alone.
Plan chat only helps with the written Next step (here: a short newsletter draft) and stops on freestyle asks.
That’s “sustainable AI interference”: the AI helps where the plan allows, and doesn’t quietly take over the rest.

Also try the step-by-step alpha demo (no live model). Real tool: local Mechanicall with a real CURRENT.md on your machine.

Show the plan the right-hand chat must follow

Loading…