The Cheapest Employee I Ever Hired Only Says Yes, No, or Which One

Plain-English guide · 7 minutes · no code required.

Jev is a small AI model from TypeSafe that answers only three kinds of question — yes-or-no, which-one, how-much — and returns a number instead of a paragraph.

Who this is for: you run a business or a creative practice, you use Claude or ChatGPT, and you don’t write code. Who should skip it: if you already run agents and want the install and the numbers, go straight to the technical post.

After reading this you’ll know which of your weekly decisions to hand to Jev first, and how to tell if it’s working.

Numbers last verified 2026-09-22 against my own log.

So what is it?

You already use Claude, ChatGPT or Gemini to write things. Good. That’s what they’re built for.

But look at what your tools actually ask all day. “Is this a refund request?” “Is this lead worth a call?” “Which of these three should I do first?” Those aren’t writing questions. They’re sorting questions.

Right now you’re paying a full writing model about two cents and five seconds for each one. Jev answers the same question for a fraction of a cent, in a blink. It can’t write. That’s the point.

What questions can it answer?

Three shapes. Once you see them, you’ll spot them everywhere.

1. “Is this…?” You give Jev a piece of text and one yes-or-no question. You get a number from 0 to 1. “Is this email asking me for a refund?” A real refund request came back 0.98. A thank-you note came back 0.01.

2. “Which one?” You give Jev the options. It picks one and tells you how sure it is. “Does this reply need to go out today, this week, or never?”

3. “How much?” You define the levels, low to high. Jev places the thing on them. “How urgent is this ticket: Cosmetic, Minor, Major, Blocking?”

Where the name comes from

TypeSafe calls Jev a “System One” model, after Daniel Kahneman’s Thinking, Fast and Slow. System One is the fast, gut-level part of your thinking. It knows an email is spam before you finish the subject line. Jev is built to be that, and only that.

Why should a business owner care?

Think about one normal week.

• Emails come in. Some are leads, some are noise, some need a reply today.
• Tasks pile up. Some are ready to hand off. Some still need you to decide something.
• Drafts go out. Some are ready. Some aren’t.
• Calls get recorded. A few minutes are worth a note. Most aren’t.

None of those are writing problems. All of them are sorting problems. A big model is the wrong tool for sorting: you pay for its ability to write a beautiful paragraph, then throw the paragraph away and keep the yes.

Jev is the cheapest attention you will ever hire.

Here’s what it changed for me. My video-note tool (WatchNotes) used to take a screenshot every fifteen seconds and ask Claude which to keep. My task board sent every job to a coding agent, ready or not. My review process re-read every finding to decide what blocked a release. Each of those was a two-cent question, asked hundreds of times.

Now Jev sits in front. It answers the small question. The big model only runs when there’s something to write or build.

What can I do with it this week?

Five uses. Every one is a question you already answer by hand.

1. Inbox triage. “Is this a lead, a client, a vendor, or noise?” One pick per email, before you open any of them. Pairs with AI for Small Business.
2. Reply timing. “Today, this week, or never?” Most of my inbox is never. I used to open all of it.
3. Hand-off readiness. “Does this task have enough detail for someone else to do it without asking me a question?” No → it stays with you. Yes → a VA, a contractor, or an agent. This is the one nobody else writes about, and it’s the one that gave me my time back.
4. Draft gating. “Is this note ready to publish, or still thinking out loud?” Plugs into Done Is a System, Not a Feeling.
5. Recording review. “Did the speaker just say something worth writing down?” Paragraph by paragraph. This is how I turned 838 hours of recorded calls into pricing and proof.

Where does it go wrong?

Three things I learned by getting them wrong.

It’s not a fact-checker. I asked whether a summary was faithful to the original. It said 0.38 about one that was perfectly faithful. Jev reads tone, words, shape. It doesn’t verify claims. Check: never gate anything on “is this true.”

It only knows what you show it. I once left two numbers out of what I gave it. It confidently ranked the wrong item first. No error, just a plausible wrong answer. Check: if a confident answer feels off, look at what you left out before you blame the model.

Low confidence means “I don’t know.” Under 50% is not “probably.” It’s no opinion. Check: below 0.5, do what you would have done anyway. I learned this when it sent a job to my cheapest AI helper at 97% and the helper stopped four-fifths of the way through. The job’s average step was easy. Its hardest step wasn’t. Now I ask about the hardest step.

How does it fit with the rest?

I run my business and my notes out of Obsidian, with Claude Code doing the heavy lifting. The agentic PKM guide is the map. Jev is one rule added to it: if the question is sort, route, pick or score, ask Jev first, a big model second, and always have a fallback if Jev is down.

The habit I’d copy even if you never use Jev: keep a log of every decision you hand to a machine. Date, where, the question, the answer, what it saved. One file. Every new tool reads it first. Same idea as Your Notes Are Not Your Knowledge, applied to automation.

Do this in 10 minutes

1. 01
   Write one question as a plain sentence. The one you answer most by hand. "Is this email asking for a refund?" is a fine first one.
2. 02
   Test it on twenty real examples. Ten that should be yes, ten that should be no. If the yeses land above 0.7 and the nos below 0.3, it's ready. If not, add what you left out of the text.
3. 03
   Write the log row. Date · where · the question · the result · what it saved. That row is worth more than the tool.

Send this to the friend who keeps asking which AI they should use. The answer for half their questions is “the one that only says yes or no.”

Questions people ask

Is Jev free? No. It’s a paid API from TypeSafe with usage pricing; my real cost is about $0.0002 a question. Current terms at typesafe.ai.

Is Jev a chatbot? No. No conversation. You send a question and some text, you get a number.

Jev vs ChatGPT or Claude: is it the same thing? No. They write. Jev decides. Use Jev to decide whether the writer needs to be called at all.

Why not use a small model like Haiku for yes/no? You can; I did. A small chat model still returns a sentence you have to parse, still takes seconds, and can still wander. Jev returns a number and can’t wander because it can’t write.

Can I use Jev without Claude? Yes. It’s independent. It works in front of ChatGPT, a spreadsheet, or a human inbox.

Do I need to code? To wire it into a tool, someone does, about fifteen lines, and I’ve published mine. If you run Claude Code, point it at the technical post and ask it to do the wiring.

Is my data used to train it? Read TypeSafe’s privacy policy before you send anything sensitive. I don’t quote it here because it changes.

Where do I get it? typesafe.ai. You’ll need an API key.

Sources and further reading

• TypeSafe AI — typesafe.ai. Their own description: “an AI lab building machine-native intelligence infrastructure for automation, designed to make decisions within software.”
• Daniel Kahneman, Thinking, Fast and Slow (2011) — the System 1 / System 2 idea the model is named after.
• My technical post — Jev in front of Claude Code: install, the nine logged uses, the fail-open code.
• The Jev Decision Card — one printable page, free with the email above.

The tools I run, in plain English

• What Is DeepSeek Harness? — the open-source Claude Code alternative and the off-peak trick.
• What Is Hermes Agent? — the self-hosted team lead that runs agents while you sleep.
• Jev in front of Claude Code — the technical build log with the numbers.
• Claude Code for Non-Coders — start here if none of this makes sense yet.