Pingo Creative TechnologyPCTBlog

← All articles

Yesiree: An Open-Source Way to Make AI Agents Ask Before They Act

By Kenneth Looi (Pingo) 3 min read

TL;DR: Yesiree is an open-source, MIT-licensed security prompt you paste into an AI agent like Claude Code, Cursor or Codex. It treats every computer access as denied until you say yes in your own words, with stronger confirmation for riskier actions. Try the live demo and its approval simulator.

A user asked an AI assistant a simple question: “How do I share my project files with you?”

Instead of explaining, the assistant started searching their computer for project folders.

Nobody asked it to look. Nobody approved it. It got as far as an approval prompt before the user stopped it, so nothing was read. But the trust was gone.

The gap between what you asked and what it decides

AI agents are getting genuinely useful. They can read files, run commands and use your accounts to get real work done. Most of the time, that’s exactly what you want.

The danger is the gap between what you asked and what the agent decides would help. A helpful agent with too much leeway is a risk to your data, your customers and your reputation.

The right answer was simple: “You can paste or attach the files, or tell me the exact folder path, and I’ll ask before opening anything.”

What Yesiree does

Yesiree is a set of plain-English rules you place in your agent’s highest-priority instructions. At its core is a three-tier approval system:

  • Tier 1, low risk: reading one specific file you named. The agent asks once: “Are you sure?”
  • Tier 2, medium risk: editing files, installing tools, running commands that change things. The agent asks a stronger question and says what could be lost.
  • Tier 3, high risk: broad searches of your machine, personal files, credentials, sending anything, using your accounts, deleting things. This needs a double yes in two separate messages, the second a typed confirmation phrase.

On top of that, it tells the agent to:

  • Treat silence, timeouts, auto-approvals and relayed requests as no.
  • Never scan your whole computer to “find” something. Ask for the exact path instead.
  • Never read, copy or send passwords, keys or tokens.
  • Report honestly. No saying “done” when the work was only queued.
  • Say plainly what an action will do, where, and what the risk is, before it runs.

One yes, one action. A new target or a wider scope needs its own confirmation.

Why a business owner should care

You might never paste a prompt into a coding agent. But if you’re bringing AI into your business, whether that’s a receptionist, an enquiry assistant, or an agent that handles your admin, the same question applies: what is it allowed to do without asking you?

Yesiree is how I think about that question, written down and made public. When PCT builds AI for a business, the same principles go in:

  • Scoped access. The AI only touches what it needs for the job.
  • Clear approvals. Anything risky, like sending, spending or deleting, waits for a human.
  • Honest reporting. You always know what actually happened, not what the AI hoped happened.

It’s also a reminder that a prompt isn’t a security wall. Yesiree’s own docs say to pair it with your tool’s permission settings, block sensitive folders, and run risky work in a sandbox. Good AI systems use layers.

Try it

Setup is one step: paste the prompt into your agent’s rules file (CLAUDE.md, .cursor/rules/, AGENTS.md and so on) and set the folder it’s allowed to work in.

Want AI in your business that knows its limits?

If you’re thinking about AI agents, assistants or automations for your business and want them built safely, with the right approvals in the right places, get in touch through the PCT contact form or email pingo@pingocreativetechnology.com. More of my work is at kennethlooi.com.