From dataset to a better prompt.
Four steps stand between your labelled examples and a prompt that scores higher on your own data. No license keys, no setup ceremony.
- 01
Install and point at a pipeline
Install the CLI, then aim PromptPotter at any backend that publishes a pipeline definition. It reads the tunable params and never edits your code.
pip install prompt-potter - 02
Bring a labelled dataset
Drop in input/output pairs — the examples every candidate prompt is scored against. A few dozen is enough to start.
datasets/labelled.jsonl - 03
Run the loop
Each round proposes a population of candidates, scores them on your data and critiques the result. The next round builds on that evidence.
prompt-potter run --pipeline ./pipe.yaml --data ./labelled.jsonl - 04
Watch the numbers move
Follow rounds in the browser. When L1 stalls, L2 reframes the task; if that stalls, L3 replans — each layer fires only on real evidence.