Prompts#

A prompt is a project item with typed variables and a version history, not a string buried in a script.

Writing one#

A prompt has a system message, a template, and variables written as {{name}}:

System:   You are a careful summariser.
Template: Summarise the following in {{sentences}} sentences, in {{style}} style.

          {{text}}

Press ＋ add … from the template and every placeholder becomes a declared variable. Each one has a type:

Give a variable a default and it becomes optional; mark it not required and it may stay empty.

Versions#

A new version is created only when the content changes — the system message, the template, the variables or the model settings. Renaming the prompt, editing its description or adding a tag does not create one.

History lists every version. Click one to see a line diff against your current draft, and Restore to bring it back into the draft.

This is what makes an experiment reproducible: it pins a version number, and that version cannot drift.

The playground#

Fill in the variables — enums get a dropdown, booleans a toggle, everything else a text box — and press Render to see exactly what would be sent, or Run to send it.

Two behaviours worth knowing:

• An empty value is treated exactly like a missing one: the default applies, and if there is no default the prompt reports it as missing and refuses to run.
• An unknown placeholder is left visible in the rendered text rather than blanked, so a typo like {{txt}} is obvious immediately.

From a script#

local text = prompts.run("summarize", { text = pageText, sentences = 3, style = "plain" })
local rendered = prompts.render("summarize", { text = pageText })
print(rendered.user)
print(table.concat(rendered.missing, ", "))

Next#

Run a prompt across a dataset and several models with an experiment.