Prompts · StartupBuilder

A file in prompts/ is one Template of model instructions, filled from records by a Role, a Function or a Workflow step. Every edit is a Version, graded before it merges.

leaseline/
prompts/
support-reply.mdthe instructions in use
support-reply@lease.mda challenger, inert until an Experiment names it

Reference

Markdown with frontmatter. The body is the instructions; {{name}} reads an input.

Key	Type	Holds
`description`	string	When it is used, in one sentence. Required
`inputs`	map	Each field it fills and its type: a primitive or a Noun name
`model`	`draft \| classify \| judge \| agent`	The task class the AI gateway routes; never a model name
`evaluators`	Evaluator ids	What grades every output

A Role file names a prompt with instructions: prompts/support-reply in place of a body. A Function names one with instructions: prompt('support-reply').

Examples

Minimal

prompts/support-reply.md

---
description: Answer a property manager's question about the product.
---
Answer in plain words. Link the page that shows how.

Leaseline

prompts/support-reply.md

---
description: Answer a property manager's question about one of their leases.
inputs: { conversation: Conversation, lease: Lease }
model: draft
evaluators: [grounded-reply]
---
You answer for Leaseline, which tracks lease renewals for property managers.
The manager asked about the lease ending {{lease.endsOn}}. Answer from that record.
Never promise a rent or a date the record does not hold. If the record cannot answer, escalate.

The challenger, support-reply@lease.md, keeps the frontmatter and rewrites only the body. An Experiment file that names varies: prompts/support-reply runs the two as arms.

What it builds

The StartupBuilder compiles each file on merge.

a Template, with one Version per edit; each Action records the Version it ran
a Variant per @arm file, assigned only while an Experiment runs
a Change check: each Evaluator runs on the Role's recorded cases, and a Version that scores worse than the one in use is refused

Good to know

A prompt is instructions, not a Permission. The Role's Mandate and Limits bound every call the model makes.
Changing the model is a Change to a Policy, never to a prompt.
A challenger runs in shadow on live inputs, with no one-way Actions, before any Customer sees its output.
Instructions an Agent loads on demand, with scripts beside them, belong in a skill package.

ReferenceCopy Anchor Link

ExamplesCopy Anchor Link

MinimalCopy Anchor Link

LeaselineCopy Anchor Link

What it buildsCopy Anchor Link