Generate, score against real data, critique, repeat. A measured search, not a guess.
It is a program optimizer for prompts and pipeline settings. Each round it writes new versions and scores them on your data. The winner becomes the next parent, and you keep the number that proves it. Evolve agents, without retraining your models.
It reads what your backend already publishes. 4 are wired today; the rest are open slots anyone can ask for.
How a connector works →Beside the question is what it runs: one call, three steps, or an agent.
One call: your input, an LLM, your answer.
Three steps: your input, a web search, an LLM, a ranking, your answer.
An agent: your task, an agent working on its own in a sandbox, a checked result.
Running something we don’t list? Tell us what it is
What is your AI getting wrong?
the drawing below runs triage-v1
$0.76 spent lift +0.21 94% from 73%
The prompt this candidate actually ran:
Read the ticket and decide which team it belongs to. Use only what the ticket says — do not guess at anything it leaves out. When the ticket describes a symptom rather than a cause, route on the symptom. A ticket naming a specific error code goes to the team that owns that code. If nothing in the ticket settles it, say so rather than choosing.
112 now right 7 now wrong vs origin 500 both measured
illustration An illustration in the shape of a real run. These figures are a worked example, not a measurement.
Every score it measures also reaches up and out to whatever you already run.
One round from a real run. Three versions were scored on the same 34 examples. The third was dropped after 7; the other 27 were never paid for.
| Version | Examples scored | Chance it is the best | Outcome |
|---|---|---|---|
| A | 34 of 34 | 50.0% | won the round |
| B | 34 of 34 | 49.7% | ran to the end |
| C | 7 of 34 | 15.8% | cut after 7 |
The whole round cost $0.0311. One round, replayed ↗
20 public benchmark cuts screened, 4 admitted. How each of these was measured →
Share of answers correct on BIG-Bench Extra Hard mini, before any optimizing.
28% starting score, admission screen
Scored before the dataset moved to a different model. The protocol says to re-measure before quoting it against anything.
Share of deductive-logic questions answered correctly, before any optimizing.
50 to 62.5% starting score, admission screen
Floor 35%. From the recon slice, not the current wiring. The protocol says to re-read it before quoting.
Share of competition maths problems answered with the exact integer, before any optimizing.
30% starting score, admission screen
Share of long-context search questions a judge grades correct, one agent episode each.
26.3 to 33% starting score, published figure
The published figure without a skill, not our own screen: a judge-graded agent episode cannot be screened at the bar. Published ceiling 39.4 to 44.7%.
A screenshot of one real, finished campaign. Every figure in it is a reading from a small run, not a benchmark result.
The app, screen by screen →
Routers send each call to the cheapest model that can still do the job, and the bill falls. Not one of them touches the thing being sent.
| What can change | A router | PromptPotter |
|---|---|---|
| which model runs each step | yes | yes |
| the wording of the request | no | yes |
| the examples you show it | no | yes |
| thresholds and retries | no | yes |
| any pipeline setting you expose | no | yes |
The two stack rather than compete: OpenRouter is the default provider here, not a rival.
Routers: Switchyard, OpenRouter, LiteLLM, Portkey, Devin Fusion, Kong, Bifrost, RouteLLM, Not Diamond.
Online, limited
10 campaigns · up to 10 rounds each
Secure, personal, local
Unlimited campaigns and rounds
Team online
Unlimited · multi-user
Try it on your own examples Built and run by one person. About →