---
title: "Three Build with Jev Recipes, Rebuilt in FlowRunner With a Person in the Loop"
description: "We gave three of TypeSafe AI's Build with Jev recipes to FlowRunner Prompt to Flow and got working flows that call Jev: a vaccination record check, an invoice field check and an AI answer check. Here are the prompts, how each flow works, how our results compare with TypeSafe's, and the Slack review step we added."
url: https://flowrunner.ai/blog/build-with-jev-recipes-in-flowrunner
date_published: 2026-11-12
date_modified: 2026-11-13T14:04:35-08:00
---

Product November 12, 2026 Updated November 13, 2026 10 min read

# Three Build with Jev Recipes, Rebuilt in FlowRunner With a Person in the Loop

We gave three of TypeSafe AI's Build with Jev recipes to FlowRunner Prompt to Flow and got working flows that call Jev: a vaccination record check, an invoice field check and an AI answer check. Here are the prompts, how each flow works, how our results compare with TypeSafe's, and the Slack review step we added.

![A FlowRunner Prompt to Flow card on a teal field with the statement Three Jev recipes, rebuilt. and the line Plain English in, live flows out.](https://flowrunner.ai/images/blog/build-with-jev-recipes-in-flowrunner-hero.webp)

[TypeSafe AI](https://typesafe.ai) publishes a set of [Build with Jev recipes](https://docs.typesafe.ai/cookbooks). Each one solves a real problem with Jev, their model that answers questions with a probability instead of a paragraph, and each one is written in Python.

We wanted to know how close we could get to those recipes without writing any code. So we picked three, described each one to [FlowRunner Prompt to Flow](https://docs.flowrunner.ai/build/prompt-to-flow) in plain English, and let it build the flows. They run in our workspace and call Jev through the [Jev integration](https://flowrunner.ai/integrations/jev). Where a recipe has Python logic, the flow has a JavaScript step that does the same thing.

Then we added the part every recipe leaves to you. When Jev isn’t sure, the recipe flags the item for review and stops. Our flows send that item to a person in Slack, and record what they decided.

## Why Jev fits a flow

Most AI steps write text, and a flow then has to guess what the text means. Jev doesn’t write text. You give it the content and a typed question, like “is this vaccine on the record?” or “where on the page is the invoice number?”, and it returns a probability or a box. The flow branches on the number.

That number also tells you when to stop and ask. A probability of 0.97 is an answer, and 0.40 is a reason to get a person involved. TypeSafe’s recipes are built around this idea, which is why they convert so well into flows with a human step.

## How we built them

Each recipe got its own chat with Prompt to Flow connected. We pasted a description of what the flow should do, the way a business user would explain it, and Prompt to Flow built the flow in our workspace: the Jev questions, the JavaScript steps that apply the recipe’s rules, the Google Sheets log and the Slack message.

Then we ran each flow on TypeSafe’s own sample files and compared our scores with the ones the recipes publish. All three flows ask Jev questions on the same pinned model the recipes use, `jev-1.14.0`, so the numbers are comparable.

The prompts are below, so you can paste them into your own chat and get the same flows.

## Recipe 1: Checking a scanned vaccination record

TypeSafe’s recipe, [Reading a scanned vaccination record](https://docs.typesafe.ai/cookbooks/reading_scanned_documents), reads a travel vaccination record that’s only a photo, with no text in it. A travel clinic wants to know, before a trip, which doses are on the record, whether each date can be read, and whether each dose will still protect the traveller.

This is what we asked for:

> Build a flow that checks a scanned travel vaccination record before a trip, using Jev from TypeSafe AI. Start it with three inputs: the scan’s image URL, the departure date and the traveller’s name.
> 
> First ask Jev, about the scan, whether each of the 16 vaccines our clinic gives is on the record (Hepatitis A, Hepatitis B, Hepatitis A and B combined, Typhoid, Hepatitis A and typhoid combined, Yellow fever, Rabies, Japanese encephalitis, Tick-borne encephalitis, Cholera, MenACWY, MenB, Tetanus, diphtheria and polio (Td/IPV), MMR, Varicella, Influenza), where its row is, and whether the table lists any vaccine outside that list. Count a vaccine as on the record at 0.5 or more.
> 
> For each vaccine found, ask Jev what year the dose was given (2016 to 2026, other, or unreadable; dates are day/month/year) and whether the date is fully legible. Then, without the image, ask Jev whether each dose will still protect the traveller on the departure date.
> 
> Flag a row “Check the reading” if finding or reading it scored below 0.90, legibility is below 0.50, or the year is other or unreadable. Flag it “Check the dose” if the in-date score is below 0.50. Flag the whole page if an unlisted vaccine scores 0.5 or more.
> 
> If anything is flagged, add a row to my “Jev review log” Google Sheet with the review id, the flagged rows and the status pending, then send me a Slack direct message listing the flagged rows with the scan and two buttons, Confirmed and Needs follow-up, and finish. When I click a button, record my choice, my name and the time in that row, and replace the buttons with what was recorded. Ignore a second click on a review that is already decided. If nothing is flagged, finish with the all-clear.
> 
> Use my Jev key and pin the model to jev-1.14.0.

The flow asks Jev three rounds of questions. The first finds each vaccine and its row on the page. The second reads the year of each dose, with “unreadable” as an allowed answer so Jev doesn’t have to guess. The third works from text only and asks whether each dose still covers the trip. A JavaScript step then applies the clinic’s rules and decides which rows need a nurse.

![The Vaccination Record Check flow on the FlowRunner canvas. Code steps list the vaccines and prepare each round, three Jev steps find the vaccines, read each year and its legibility, and check each dose covers the trip, then a code step flags rows for the nurse. If anything is flagged, the flow logs the review as pending in Google Sheets and asks the nurse in Slack; if not, it finishes with all clear.](https://flowrunner.ai/images/blog/build-with-jev/flow-vaccination-record-check.webp)

**How it compared.** The flow flagged the same two rows as TypeSafe’s recipe. The Rabies date is under a batch sticker, so the flow flagged it “Check the reading”, with a legibility score between 0.05 and 0.08 across our runs (the recipe shows 0.05). The Typhoid dose is from 2021 and won’t cover a trip in November 2026, so the flow flagged it “Check the dose”, with an in-date score between 0.31 and 0.36 (the recipe shows 0.38). The other four rows came through clear.

This is the message the nurse gets:

![A Slack direct message titled Please review a vaccination record, saying Sample traveller leaves on 2026-11-12 and Jev flagged two rows: row 3, Typhoid, year 2021, check the dose, and row 6, Rabies, year unreadable, check the reading. The scanned record is attached, with Confirmed and Needs follow-up buttons below it.](https://flowrunner.ai/images/blog/build-with-jev/vaccination-review-slack.webp)

## Recipe 2: Checking invoice values against the photo

The second recipe, [OCR confidence and bounding boxes](https://docs.typesafe.ai/cookbooks/ocr_confidence_bounding_boxes), starts after an invoice reader has already pulled the values off a supplier invoice photo. Readers like that sometimes get a single character wrong, and an invoice number or a bank detail with one wrong digit is the kind of mistake that costs money. The recipe checks every value against the photo twice: once against the whole page, and once against a close-up of the spot where the value is printed. If either check scores below 0.25, the value goes to review.

This is what we asked for:

> Our invoice reader pulls fields out of supplier invoice photos, but sometimes it misreads a digit. Before anything gets entered, check every value against the photo with Jev. For each field (supplier name, invoice number, invoice date, amount due, bank name, account holder, account number, branch code), ask Jev whether that value is really what’s printed on the invoice and where on the page it is, then zoom in on it and ask Jev again whether the close-up clearly shows that value. If either score for a field is below 0.25, add a row to my “Jev review log” Google Sheet with the review id, the flagged fields and the status pending, then send me a Slack message that starts with the invoice number and the supplier and shows each flagged field, the value we read, both scores and a close-up of that spot, plus two buttons: “Values are right” and “Send back for correction”. Don’t wait for me. When I click a button, record my choice, my name and the time in that row, and replace the buttons with what was recorded. Ignore a second click on a review that is already decided. Use my Jev key and pin the model to jev-1.14.0.

The flow asks Jev about every value on the full page first. Then it goes through the values one at a time: Jev finds the value on the page, the flow crops that spot, and Jev checks the crop. Whoever reviews the invoice in Slack sees each flagged value next to the close-up it was checked against, so they can decide without opening the photo.

![The Invoice Field Check flow on the FlowRunner canvas. A Jev step checks each value on the page, a code step zooms in around each value, and a loop named For each value, locate and crop with Jev runs once per value. A code step decides which values need review, and if any value is below 0.25 the flow logs the review as pending in Google Sheets and asks AP to review on Slack.](https://flowrunner.ai/images/blog/build-with-jev/flow-invoice-field-check.webp)

Inside the loop, Jev finds the value in its zoomed view, a code step crops it, and a second Jev step checks the crop:

![The inside of the For each value loop. Jev locates the value in its zoomed view, a code step crops the value, Jev checks the cropped value, and code steps add the value to the results and save the results list.](https://flowrunner.ai/images/blog/build-with-jev/flow-invoice-field-check-loop.webp)

**How it compared.** TypeSafe’s two sample invoices each have a misread invoice number, and the flow caught both. On the Milkwood invoice, the reader got AN400074 where the invoice says IN400074. The page check passed it at 0.85, and the close-up caught it at 0.10 and 0.18 across our runs (the recipe shows 0.18). On the Siyakhula invoice, INV03070 should be INV03079, and both checks failed it, at 0.02 and 0.04, the same scores the recipe shows.

The flow missed one. The Siyakhula date was read as 16/09/2028 instead of 16/09/2026. The recipe caught it on the page check at 0.17, and our page check scored it 0.50, so it passed. On one run the flow also flagged a bank name that was correct, with a close-up score of 0.17. On the next run the same value scored 0.31 and passed.

Jev’s scores move a little from run to run, and values near the line can land on either side of it. That’s the reason the flow sends uncertain values to a person instead of entering them, and why it’s worth setting the line with your own invoices in front of you.

## Recipe 3: Checking an AI model’s answer before you use it

The third recipe, [SDE cascade](https://docs.typesafe.ai/cookbooks/sde_cascade), is about cost. A cheap AI model answers a meeting-scheduling task as JSON. Jev checks the answer against six rules, such as “does the meeting overlap a time when someone is busy?”. If a rule fires, the task goes to a stronger, more expensive model. On TypeSafe’s test of 1,000 tasks, this got 94.9% of answers right for $0.26 per 1,000 tasks, against 98.1% for $1.89 when the strong model answers everything.

This is what we asked for:

> Build a flow that checks an AI model’s answer before we use it.
> 
> It starts when it is called with a meeting-scheduling task as text.
> 
> 1.  Have a cheap model (Claude Haiku 4.5) answer the task. Put this right before the “SOLUTION:” line of the task: “Return ONLY a JSON object matching this JSON Schema:” with a schema for day (Monday to Friday), start\_time and end\_time (24-hour HH:MM). Nothing else may be in the object.
> 2.  Check that the reply is valid JSON that matches the schema.
> 3.  If it does, ask Jev (model jev-1.14.0, my Jev key) six yes/no questions about the task and the answer, where yes means “the answer breaks this rule”:
>     -   Does the meeting time overlap a time when any participant is busy?
>     -   Is it outside the work hours or the days the task allows?
>     -   Does it break a preference stated in the task?
>     -   Does the task ask for the earliest availability, and is there an earlier time that works for everyone?
>     -   Is the meeting a different length from the one the task asks for?
>     -   Does the answer fail to propose exactly one day and one start and end time?
> 4.  If the reply is not valid JSON, or any rule scores above 0.7, give the same task to a stronger model (Claude Opus 5) and check its answer the same way.
> 5.  If the stronger answer also fails, add a row to my “Jev review log” Google Sheet with the review id, both answers, the rules that failed and the status pending. Then send me a Slack DM with the task, both answers and the rules that failed, with buttons to use the strong answer, use the cheap answer or reject both, and finish.
> 6.  When I click a button, record my choice, my name and the time in that row, and replace the buttons with what was recorded. Ignore a second click on a review that is already decided.
> 7.  Otherwise return the answer that passed and which model gave it.

TypeSafe’s recipe uses two [OpenAI](https://flowrunner.ai/integrations/openai-ai) models. FlowRunner runs on your own API keys from any provider, so we used Claude Haiku 4.5 as the cheap model and Claude Opus 5 as the strong one.

![The Check an AI Answer flow on the FlowRunner canvas. Claude Haiku 4.5 answers first, a code step checks the answer's format, and Jev checks the cheap answer against the rules. If the JSON is invalid or a rule fires above 0.7, Claude Opus 5 answers and Jev checks that answer too. If the strong answer also fails, the flow logs the review as pending in Google Sheets and asks a person on Slack.](https://flowrunner.ai/images/blog/build-with-jev/flow-check-ai-answer.webp)

**How it compared.** On the recipe’s example task, a one-hour meeting for Joan, Theresa and Shirley, TypeSafe’s cheap model answered 10:30, which overlaps two meetings. Jev scored that rule at 0.97 and the task went up to the strong model, which answered 14:00. In our flow, Claude Haiku 4.5 answered 14:00 on the first try. Every rule scored 0.26 or lower, so nothing was escalated and the expensive model never ran.

To see the whole path, we gave the flow a task with no valid answer. Dana and Luis need an hour together, and their calendars never leave one. The task text still says a solution exists, so both models made one up. Claude Haiku 4.5 answered 13:30, and Jev scored the busy-time rule at 0.97. Claude Opus 5 answered 12:30, and Jev scored it 0.84. Both answers looked fine as JSON, both were wrong, and the flow sent them to a person instead of handing either one on.

## The step we added: a person decides

All three recipes end with something like “flag it for review”. In a real business, someone has to see that flag and act on it. That’s the step we built into each flow.

When a check fails, the flow adds a row to a [Google Sheets](https://flowrunner.ai/integrations/google-sheets) review log with the status “pending”, sends a [Slack](https://flowrunner.ai/integrations/slack) message with buttons, and finishes. Nothing sits waiting in the meantime. When someone clicks a button, a second flow picks up the click, finds the review in the log, and records the decision, who made it and when. Then it replaces the buttons on the message with the decision, so anyone looking at the channel can see the item is handled:

![Two Slack messages after review. The first reads Invoice AN400074 from MILKWOOD mushrooms: invoice number flagged, with Values are right by markpiller at 2026-11-12 01:56 UTC, logged in the Jev review log. The second reads Vaccination record for Sample traveller, leaving 2026-11-12: 2 of 6 rows need a nurse, with Record confirmed by markpiller at 2026-11-12 02:04 UTC.](https://flowrunner.ai/images/blog/build-with-jev/decided-reviews-slack.webp)

If two people click the same message, or someone clicks an old one, the second click changes nothing, and a short note in the thread says who already decided. One click flow handles all three review types, so adding a fourth flow later doesn’t mean building another one.

![The Jev review decisions flow on the FlowRunner canvas. It starts when a reviewer clicks a Jev review button in Slack, reads the click and finds the review in the Google Sheets log. If the review is still pending, it records the decision in the log and shows the decision on the message. If it was already decided, it replies in the thread, and if the review isn't in the log, it replies that it wasn't found.](https://flowrunner.ai/images/blog/build-with-jev/flow-review-decisions.webp)

We call this [human-in-the-loop](https://flowrunner.ai/concepts/human-in-the-loop): Jev handles the clear cases on its own and asks a person about the rest. Slack is one option. The same review can go out by email, [WhatsApp](https://flowrunner.ai/integrations/whatsapp) or phone.

## What testing caught

The first versions didn’t all work on the first run. Each fix was one more message in the same chat:

-   **The cheap model asked for a task.** TypeSafe’s tasks say “Here are a few example tasks and solutions:” before the real one, and Claude Haiku 4.5 took that literally and asked us which meeting to schedule. We had the flow reword that one line before sending the task, and tell both models to reply with JSON only.
-   **The invoice messages looked the same.** Both invoice reviews started with the same sentence, so it was hard to tell which one you were clicking. They now start with the invoice number and the supplier.
-   **A click didn’t show.** The first click flow replied in the message’s thread, which Slack shows as a small “1 reply” link. Replacing the buttons with the decision made it obvious the click went through.

## Import the finished flows

If you’d rather start from what we built than from the prompts, download the flows and [import them](https://docs.flowrunner.ai/manage/flows#importing-a-flow). In FlowRunner, click **+** beside Flows, choose the file under **Import a Flow Version**, and click **Create**.

-   [Vaccination Record Check (Jev)](https://flowrunner.ai/downloads/prompt-to-flow/jev-vaccination-record-check.json)
-   [Invoice Field Check with Jev](https://flowrunner.ai/downloads/prompt-to-flow/jev-invoice-field-check.json)
-   [Jev: Check an AI Answer](https://flowrunner.ai/downloads/prompt-to-flow/jev-check-ai-answer.json)
-   [Jev review decisions](https://flowrunner.ai/downloads/prompt-to-flow/jev-review-decisions.json), the flow that handles the Slack clicks for all three

The files have our connections, keys, spreadsheet and Slack user taken out. After importing:

1.  Paste your TypeSafe key into Jev’s settings in your workspace, if you haven’t already.
2.  Connect Slack and Google Sheets when FlowRunner asks. Use the same Slack account in all four flows, so the review decisions flow can update the messages the other three send.
3.  Make a Google Sheet called “Jev review log” with these columns in row 1: Review ID, Flow, Run ID, Created At, Summary, Payload, Status, Decision, Decided By, Decided At. Pick it on each Sheets step.
4.  On each Slack step that sends a review, pick the person who gets it.
5.  In the AI answer flow, pick your own AI key on the two model steps.

## Build these yourself

1.  **Get a TypeSafe API key** at [console.typesafe.ai](https://console.typesafe.ai/keys) and paste it into Jev’s settings in your FlowRunner workspace.
2.  **Connect Prompt to Flow** to the AI tool you already use, such as Claude, ChatGPT or Cursor. The [setup steps are in the docs](https://docs.flowrunner.ai/build/prompt-to-flow).
3.  **Paste one of the prompts above** into a new chat, with your own Slack and Google Sheets connected. Swap in your own rules: your clinic’s vaccines, your invoice fields, your models.

If you don’t have a FlowRunner account yet, [start free](https://app.flowrunner.ai). No credit card, and every integration, Jev included, is on every plan.

## Need it built for you? Our team does custom work.

Tell us the process. Our implementation team at Midnight Flow builds it around how you work, tests it on your real data, and keeps it running, in your own FlowRunner workspace.

[Have us build it](https://midnightflow.ai/packages/custom?utm_source=flowrunner.ai&utm_medium=referral&utm_campaign=general&utm_content=blog.aside&utm_term=build-with-jev-recipes-in-flowrunner)

## Frequently asked questions

### What is Jev?

Jev is a model from TypeSafe AI that answers typed questions about text, JSON or images and returns a probability instead of written text. You ask whether something is true, which option fits, or where something is on a page, and the flow branches on the number that comes back.

### Do I need to write Python or JavaScript to build these flows?

No. We described each recipe to FlowRunner Prompt to Flow in plain English, and it built the flow, including the JavaScript steps that replace the recipe's Python. You can open and edit those steps, but you don't have to.

### Can I use a different AI provider for the cheap and strong models?

Yes. FlowRunner uses your own API keys. TypeSafe's recipe uses OpenAI models; we used Claude Haiku 4.5 and Claude Opus 5 from Anthropic. Any provider you have a key for works the same way.

### Can someone build these flows for us?

Yes. Our implementation team at Midnight Flow builds flows like these around your own documents, rules and review process, in your own FlowRunner workspace. Use the Have us build it button on this page to tell us what you need.

By [Mark Piller](https://flowrunner.ai/about), Founder of FlowRunner [Editorial policy](https://flowrunner.ai/editorial-policy)

---
Markdown version of https://flowrunner.ai/blog/build-with-jev-recipes-in-flowrunner. Site index: https://flowrunner.ai/llms.txt
