The Prompt Registry · Early Access

Ship prompts
like code

Author, version, and publish LLM prompts as packages. Install with the CLI and get type-safe generated code — no copy-paste, no guesswork.

Get StartedStar on GitHub
Your prompts are product surface area. Treat them like it.
How it works

Three steps from
prompt to production

STEP 01
Author
Write templates with typed Mustache variables. Add docs, output schemas, and metadata.
Hi {{user.name}}{{@type string}} Welcome to {{productName}} {{@outputSchema}}
STEP 02
Configure
Declare prompts with version ranges in sufleur.yaml. The CLI resolves and locks them.
# sufleur.yaml prompts: '@acme/welcome': '^2.0.0' output: typescript
STEP 03
Generate
The CLI fetches templates and generates type-safe code your app imports directly.
$ sufleur add @acme/welcome $ sufleur generate ✓ wrote prompts.ts
The gap

From string concatenation
to a real dependency

app/services/email.tsBefore
// 😬 a string, somewhere in your app const p = `Hello ${name}, welcome to ${product}.` // no types · no version · no validation const res = await llm.chat({ messages: [{ content: p ]} }) const data = JSON.parse(res) // 💥 fingers crossed
Features

Everything you need to
ship prompts like a package

Type-safe codegen
The CLI generates TypeScript & Python with full type inference — no manual types, no runtime surprises.
const { prompt } = welcome.render({ name: 'Ada' // string ✓ })
Draft → Published lifecycle
Iterate on drafts without touching production. Publish when ready — only published versions are installable.
CLI-first DX
One binary, zero config. sufleur add & generate — that's the whole workflow.
Semver versioning
Immutable versions, pinned or ranged — just like npm.
Output schemas
Zod & Pydantic, generated.
Workspaces
Roles for authors & consumers.
Evals

Ship every version
with receipts

Define an eval once — a dataset, checks, and an optional LLM judge. Every version gets graded against it, and the at-publish result is frozen onto the prompt's public page.

Every run pulls @acme/[email protected] — 8 versioned cases with inputs and expectations, exposed to checks as case.*.
Sample run

Every run produces a report like this

run_01hq6r@acme/support-triage@2.3.0succeededpassed
88%
cases pass
verdict: passed · threshold 80%
7/8
cases passed
21/24
checks passed
19s
run duration
claude-haiku-4-5
model
severity is valid88
valid output schema88
routing quality ≥ 788
8 cases · 3 assertions · 1 judge · 24 checks — click a case to inspect it
Decision models

Prompts for TypeSafe Jev,
versioned like the rest

System-One decision models like TypeSafe’s Jev don’t write text — they answer typed questions with calibrated probabilities. Sufleur gives your Jev prompts the same registry, playground, evals and codegen as your LLM prompts.

Versioned questions & rubrics
Instructions, choice options and score levels are semver’d. Adding an option is a breaking change, and the diff says so.
Typed answers in your code
getDecision() gives you a typed batch: ask any question as often as your data needs, then read back answers whose choices are literal unions in Zod & Pydantic.
Evals on every version
Grade each version against a dataset with CEL checks like output.department.choice == case.team.
Decision model docs
@acme/ticket-triagejev-latest · 3 questions
isRelevantnoul
Is this ticket about our product?
0.94
departmentchoice
Which team should handle ticket.subject?
billing
81%
technical
12%
sales
7%
frustrationscore
How frustrated is the customer in ticket.body?
Calm
Frustrated
Very angry
AI Agents

Hand the loop to
your coding agent

The CLI ships a skill your agent can load — so it can author, version, and publish prompts on your behalf, end to end.

Ships as a skill
sufleur skill prints a drop-in skill file for Claude Code, Cursor, and any agent.
Drives the full loop
Draft versions, edit files, set schemas, edit and run evals, render locally to test — all via the CLI.
Machine-readable
Every command supports --json, so agents act on structured output.
Cclaude code+ sufleur skill
you  Add a refund-classifier prompt → category + confidence
↳ loaded sufleur skill
$ sufleur prompt create @acme/refund-classifier
$ sufleur version draft @acme/refund-classifier
$ sufleur file update classify.mustache
$ sufleur eval run @acme/refund-classifier
✓ eval passed — 8/8 cases
✓ drafted v0.1.0 — ready for review
Get started

Stop copy-pasting prompts.
Start shipping them.

Create your workspaceRead the docs
$ sufleur add @your-workspace/first-prompt