Quick Answer
Build a chain of steps — prompt steps (run by an AI model) and transform steps (pure browser text ops: trim/join, extract regex, replace regex) — where each step’s output pipes into the next via {{prev}} (or {{input}} for the starting text). Press Run chain: steps execute in order with each result shown and its cost accrued; chains save to your library as JSON. The default example extracts action items from meeting notes and drafts the follow-up email. Free trial covers 3 model calls a day; unlimited chains on your own key.
How It Works
Enter the starting text (meeting notes, a document, anything) — it reaches prompt steps as {{input}}
Build the chain: prompt steps run on a model, transform steps (join, extract regex, replace regex) reshape text for free; {{prev}} pipes the previous output forward
Press Run chain — each step’s result and cost appears as it completes, and you can stop after the current step
Save the chain to your library or copy the final output
Key Facts
- Prompt steps + free client-side transforms (join, extract regex, replace regex)
- {{input}} receives the starting text; {{prev}} receives the previous step’s output
- Per-step results and cost shown as the chain runs; stop after the current step any time
- Chains save to the Prompt Library and export as JSON
- 3 free model calls a day; unlimited on your own API key
Frequently Asked Questions
What are transform steps?
Free steps that reshape text in your browser without a model call: trim/join cleans whitespace, extract pulls every regex capture group as a line, replace runs a regex substitution. Use them between prompt steps for deterministic cleanup — they cost nothing.
How does piping work?
The chain carries one running text. {{input}} in a prompt refers to the starting input; {{prev}} refers to the previous step’s output (identical for step 1). Transforms always operate on the previous output directly.
Can steps use different models?
Steps run on your active provider/model from the key panel. For comparing a step across models, run the chain once per model — the run log records each run’s cost and latency for comparison.