Build things that get better on their own, and never change behind your back.
Give this to Claude at the start of anything you build: a dashboard, a Shopify app, a mechanism inside a website, or an agent that just writes reports. It makes the thing self-improving, and it makes every improvement visible to you before it can change anything that matters. Automation proposes; a person disposes.
Decide
You approve or decline. Anything irreversible passes two gates: a plain-English confirm that states the real consequence, and a server-side check that refuses unless the state is exactly right. Only trivially reversible changes you pre-authorized skip the click, and even those log themselves instantly.
The cycle visibly pauses here for you. The system runs autonomously right up to the edge of anything that matters, then stops and asks, and it captures your reason when you say no.
The system may propose any improvement at any time. It may only apply a change after it is written to a human-readable change ledger, and after either you approved it or it is a narrow class of reversible changes you pre-authorized and it is logged the instant it applies.
Nothing improves invisibly. If the system cannot record a change, it cannot make the change. Everything else is the machinery that makes that sentence true and legible.
Six parts turn the rule into something real
Each is a working piece of the live system this standard was extracted from. Skip one and it stops being a loop: it becomes a black box, or a report nobody acts on.
The rules that keep a self-improving system honest
The hard-won ones: the difference between a loop that compounds in your favor and one that quietly drifts somewhere you never agreed to go.
Same principles, smaller machinery
Never drop the iron rule or the visible report, no matter how small the thing is. What shrinks is the implementation, not the idea.
Paste this into the first prompt of any build
It tells Claude to build to this standard and to ask you the right setup questions before writing a line.