Claude Skill
The Decision Check
One workflow from your product manager operating system. Test a sorting question the way you'd test an experiment: label 20 examples, write the bar down first, then run it.
Want the skill?
Install & use
One question, 20 examples you label yourself, and a pass bar written down before anything runs. Then the catches, the false flags and every miss, against that bar.
- 1
Download the file. Don't unzip it.
- 2
Open Claude Code and start a new session.
Install this skill for me. Then tell me, in plain English, what it does and what I should say to start using it. - 3
Drag the file straight into the message box.
- 4
Paste the line below, and press enter. Claude does the rest.
That is the whole install on Claude Code and Cowork. In a browser or the desktop app Claude can't reach your files, so use Settings, Customize, Skills, Add, Upload skill instead.
How it works
A marker for good enough
You set success criteria before an experiment. The same discipline rarely reaches sorting: which tickets are urgent, which requests are new. That sorting is more and more often handed to a model with no agreed marker for good enough.
This skill gives you one: a question, 20 examples you label yourself, and a bar written down before anything runs.
Use it
Claude goes one step at a time and waits for you at each.
- 1.Name the decision the answer will drive.
- 2.Write one yes/no or pick-one question.
- 3.Label 20 examples yourself.
- 4.Write the bar down, then run it.
What comes back
From the made-up example in the download, with real output:
| Blockers caught | 100% (10/10) | at least 80% | yes |
| False flags | 0% (0/14) | at most 20% | yes |
Closest call: 0.56 on "My accountant has no way to see
the books without my password." Test more like it next.Optional: bigger batches with Jev
With no key, the skill works in your Claude chat for up to about 150 rows. For thousands of rows, or a check you run every month, add your own OpenRouter key and the skill sends each row to Jev, a decision model from TypeSafe that answers with a probability instead of writing text.
Every row you run that way goes to OpenRouter and on to TypeSafe, on an endpoint marked alpha, with a proprietary model. In our test it cost about 2 to 3 US cents per 1,000 rows. The skill shows you the estimate and sends nothing until you say yes. Check your company's data policy before you send customer text.
Questions people ask
- What is The Decision Check?
- The Decision Check is a free Claude skill and one workflow of a product manager's operating system. It helps a product manager write one sorting question, label about 20 examples, set a pass bar before running, and then measure the catches, false flags and misses against that bar.
- Why write the bar down before the run?
- A bar chosen after seeing the results tends to fit those results, so it proves nothing. Writing it first turns the run into a real test. If you want a different bar, you write a new one and run again.
- Is my data sent anywhere?
- Not unless you add a key. With no key, the skill runs inside your own Claude conversation, under the same data settings as the rest of your Claude account. With your own OpenRouter key, every row you choose to run is sent to OpenRouter and on to TypeSafe, who run Jev, on an endpoint marked alpha with a proprietary model. The skill shows the row count and cost first and sends nothing until you agree. Check your company's policy before sending customer text.
- What does it cost?
- The skill is free. With no key there is nothing else to pay. With your own OpenRouter key you pay OpenRouter directly for Jev: in our test that came to about 2 to 3 US cents per 1,000 rows, more if your context is long. The skill prints an estimate before it sends anything.
- Is The Decision Check free?
- Yes. The download asks for your name and email, and you also get the weekly Product System newsletter, which you can unsubscribe from at any time.