Twenty minutes and about three cents. The example is support-ticket triage: a message comes in, one prompt summarizes it, a second picks the queue.
The Library on the projects page holds curated templates — a program, its settings and a sample dataset, ready to run. Adding one gives you a project of your own; changing it changes nothing for anyone else. Ticket triage is the one this page describes. If you would rather build it yourself, the empty canvas opens with a twelve-card tutorial that walks the same path and will do each step for you if you ask it to.
Drag Prompt from the palette twice. Click the first, name it summarize, and give it a template. Words in {braces} are placeholders filled from your dataset's columns:
Read this customer message and state in one short line what the
customer wants and which product or order it concerns.
Message: {message}Add an output field named summary. Prompts return structured fields rather than loose text, so the next prompt and the scorer can pick one out. Then the second prompt, named route, with an output field queue:
Assign this support request to one queue: billing, shipping,
refund, bug, account, or other.
Request: {summary}Drag from the right dot of summarize to the left dot of route. Click the new edge and map {summary} to the summary field. Every prompt can also read the dataset columns directly, so {message} would work in either one.
The prompt with no outgoing edge is the one Evaluate grades — here route. Click the Evaluate block and point an exact-match scorer at the output field queue. More on scorers →
On the Data tab, load the sample: 50 tickets, each a message and the queue it belongs in. Map {message} to the message column and choose queue as the label. Check data validates the wiring. More on datasets →
The pilot runs your prompts as written on twelve rows, twice, plus one rewrite. That gives you three numbers before you commit: the baseline score, the noise between two identical evaluations, and the projected cost of the full run. If the baseline is already near the ceiling, or the noise is bigger than the improvement you are hoping for, stop here — you have learned something for a fraction of a cent.
Two rounds is plenty for a first look. Watch the best-so-far curve and the tree; click any node to read its prompt diffed against its parent. When you like one, Publish it as a version. How to read a run → · Publishing and serving →
A run of this shape — 50 rows, two prompts, ten rounds — is roughly a thousand billed model calls, about a cent and a half on small models. The pilot is a fraction of that, and it runs first for exactly that reason. The projected cost is shown before you start, and every run carries a hard dollar budget that stops it.