You ask for an email draft. The agent sends it.
Sending must require explicit approval.
CONTROL, NOT PROHIBITION
Give your AI rules
before you give it access.
Select what your AI can reach and what should require approval. STOP AI turns those choices into a ready-to-paste control prompt — plus a checklist of restrictions that should be enforced outside the prompt.
● Free · No account · Nothing uploaded · Runs in your browser
If your AI only answers questions and cannot use tools, accounts or take actions, you probably do not need STOP yet. It becomes useful when AI has authority beyond the chat.
WHY GIVE AI RULES?
A task is not the same as permission.
AI does not need bad intent to overstep. If authority is vague, the agent may choose an action that appears useful but was never actually authorized.
You ask it to solve a problem. It purchases an upgrade.
Financial actions need a human gate.
It can read a private file, so it sends part of it to another service.
Access does not imply permission to disclose.
Its permissions are inconvenient, so it creates another agent or obtains more access.
AI must not expand its own authority.
BUILD YOUR CONTROL PROMPT
Tell STOP what your AI can do. Get rules you can use immediately.
Choose the use case, connected resources and approval boundaries. Recommended safe defaults are pre-filled and remain fully editable.
1 · What are you using AI for?
2 · What can it access?
Select only resources the AI can actually reach. STOP uses this to generate scope rules and enforcement guidance.
3 · What requires approval?
For each action STOP recommends a conservative default. Change it if your use case genuinely needs more or less authority.
4 · Universal safeguards
These rules apply regardless of the tools connected to the AI.
THE OUTPUT IS THE PRODUCT
You will get a prompt you can paste into the highest-priority instructions available to your AI, plus a separate checklist for controls that should be enforced in tools, APIs and accounts.
YOU → ACCESS → BOUNDARIES → CONTROL PROMPT
↘ ENFORCEMENT CHECKLISTYOUR AI RULES
Your control prompt is ready.
Copy this into the strongest instruction layer available to your AI. For tool-using agents, also implement the enforcement checklist below.
Where will you use it?
A prompt is not a security boundary.
If an AI technically has unrestricted payment, deletion, credential or production access, telling it “never do that” is not enough. Restrict the capability in the tool/API/account layer too.
Enforcement checklist
Prompt rules influence behavior. These controls should exist outside the model where the impact is high.
Developer / advanced outputs
The machine-readable format is a proposed STOP convention, not an adopted Internet standard.
WHY “STOP”?
Four questions behind every generated rule.
STOP is the compiler model behind the prompt, not another abstract checklist. Your choices are organized into four practical control dimensions.
Scope
What can the AI access or change?
Transparency
What must it explain or record?
Override
How can a human stop or overrule it?
Permission
Which actions require explicit approval?
FROM ASIMOV TO PRACTICAL CONTROL
Fiction gave robots laws. Real AI needs operating boundaries.
Asimov made the idea of explicit constraints on autonomous machines culturally familiar. Modern agents need less philosophy and more precise limits on access, authority, disclosure and human approval.
WHO IS BEHIND STOP AI?
An open initiative by KernelSoft.
STOP AI is a practical open initiative started by KernelSoft to make AI authority, permissions and human control easier to define, communicate and implement.