The console

You can watch the whole thing, or none of it.

A run works the same whether or not anyone is looking. But when you do look, this is what you see: what it is trying right now, what each attempt scored, and what it has cost so far. Four screens, one of which is just your own folder.

These are sketches, not screenshots — drawn from the real app, labelled so nobody mistakes one for a photo of it.

Where a run starts.

You say what you are trying to get right and attach examples where you already know the answer. That is the entire setup. From there the conversation keeps going while it works.

It’s stuck at 73%.

Send me examples where you know the right answer.

tickets.csv · 500 rows

Starting. Every round shows up here.

Happening now

live scoring round 4
round 3 — 0.86 ▲ +0.07
round 2 — 0.79 ▲ +0.06
round 1 — 0.73

This run

+31% over where you started $0.42

78 now right 6 now wrong 416 unchanged

Chat is where a run starts and where you watch it. You describe what you want, attach your examples, and every round reports back in the thread as it is scored. The card on the right is the honest summary: what the run gained, what it cost, and which answers flipped the wrong way as well as the right way.

Where you watch it.

Every attempt is on screen with the score it earned, so “it got better” is never something you have to take on trust. You can open any single attempt and see exactly which examples it got right and which it still misses.

Best so far round 5 running

0.91 ▲ +0.18 since round 1

What each round goes through

read the task
plan
write versions
score
learn

Versions tried this round

step-by-step 0.91 kept
worked example 0.88
plain & short 0.71
add an example 0.64

Dashboard is the whole run on one screen: how the score has moved, which step it is on right now, and every version it tried with the score each one earned. Nothing here is computed in the browser. Every number you see was measured by the run itself, so what is on screen and what is on disk cannot disagree.

Where you steer it.

You are not locked out while it runs. Pause it, skip a round, or take a promising attempt and branch a new run from that exact point to try something different.

round 5 of 20 spent $0.42 look ahead 2×

The run controls hover at the bottom of every screen while something is running. Pause it, skip a round that is clearly going nowhere, or spend faster to get an answer sooner. The number next to them is what it has cost so far, live, so the run can never quietly outspend what you meant to give it.

Picked up where it left off

Run 3 started from run 2’s best version rather than from scratch, so it inherited everything run 2 had already paid to learn.

Same job, different attempts

The sidebar keeps every run you have ever started, grouped by the job it belongs to. Runs branch: if one gets somewhere promising you can split off from that exact point and try a different direction, and the list shows you where each branch came from. Nothing is thrown away, so a dead end costs you the tokens it burned and nothing else.

Where the results actually live.

On your disk, as ordinary files, from the first round onward. The console is a nicer way to read them, not the only way — which is the part that matters if you ever stop using it.

campaign-04
summary.md
what it settled on
rounds
round 1
round 2
round 3
hardest cases

What it settled on

Work through the ticket one step at a time. Name the policy you are applying before you apply it. If two policies could apply, say so rather than picking one.

scored 0.91 started at 0.73 ▲ +31%

Why it kept this one

Files is the run’s own folder, browsable. Everything the console shows you lives here first as ordinary files on disk, and the console is an easier way to read them rather than a place they are hidden. Which is also why nothing is ever locked in: the results are already yours, in a format you can open without us.

Not there yet

What the console still doesn’t do.

No one-click export. You can read and copy everything, and the files are yours on disk, but there is no download button for the finished version yet.
Search is not wired up. The box is there in the app and does nothing. It will be removed or finished, not left ambiguous.
Some toggles are switched off on purpose. A few options in the chat panel are visible but disabled. They are marked as such in the app rather than quietly failing.

Try it — it’s free Where it fits