01
Create
Start structured instead of starting blank.
Choose one of nine prompt categories, use a curated template, import a portable Rylvo prompt, or create a custom prompt for a bot or workspace.
Prompts
Write prompts. Test responses. Control releases.
Create structured agent prompts, run them against a real model, compare every version, release deliberately, and connect production outcomes back to the exact instructions that produced them.
9prompt categories
8starter variables
5optimization strategies
3release environments
Prompt management workflow
Write the instruction. Control everything after it.
Rylvo Prompts is a working release system for agent instructions: structured authoring, real model testing, version history, controlled promotion, runtime attribution, and optional optimization in one place.
Prompt release workflow
Author deliberately · test with your key · keep the lineage
Draft
author · inherit · variables
Playground
resolve · run · inspect
Version
diff · note · compare
Production
promote · trace · measure
Draft prompts stay out of runtime
Active versions are recorded on traces
Rollback keeps newer history intact
02
Compose
Make prompts reusable and context-aware.
Add typed placeholders and defaults, inherit a parent with {{parent}}, scope by workflow stage, and assign to one bot or share across an agent group.
03
Test
See a real model response before relying on it.
Fill test variable values, send a user message, and inspect both the real BYOK model response and the fully resolved system prompt. Playground runs are not saved.
04
Version
Keep the source, author, reason, and result.
Every content save creates a numbered version with a change note. Compare any two versions line by line and promote an earlier version to roll back.
05
Release
Know which version is live.
Draft, Active, and Archived status controls prompt eligibility. Track version pointers in Dev and Staging, then promote a chosen version to Production when it should become live.
06
Measure
Connect prompt changes to real outcomes.
Review production or test traces, latency, version scores, optimization outcomes, and prompt-attributed model cost without mixing playground traffic into live results.
Structured prompt workspace
More than a large text box.
Organize prompts by runtime responsibility, reuse variables, inherit shared policy, constrain prompts to workflow stages, and see how bots, groups, and parent prompts depend on them.
PROMPTS / SUPPORT RESPONSE
ACTIVE · V12SourceRuntime preview9 VARIABLES · 684 TOKENS
You are {{company_name}}'s support assistant.
## Goal
Resolve the user's question using approved knowledge.
## Response policy
- Reply in {{language}} with a {{tone}} tone.
- Cite the relevant policy when available.
- Escalate if confidence is below {{threshold}}.
{{parent}}
{{company_name}}
Northstar
{{language}}
English
{{tone}}
Concise
{{threshold}}
0.72
Save versionPreview
Nine runtime categories
System, Response Composer, Stage Classifier, Action Selector, Escalation Classifier, Session Summarizer, Verifier, Retrieval, and Custom.
Variables with defaults
Use eight starter variables or define your own required and optional placeholders.
Parent inheritance
Insert {{parent}} to reuse a base prompt instead of copying policy into every child.
Bot, group, and stage scope
Attach to one bot, share across an agent group, or apply only during a matching workflow stage.
Dependency graph
See prompt-to-bot, prompt-to-group, org-level, and parent-child relationships visually.
Portable JSON
Export a prompt with version history and import it into another workspace subject to quota.
SystemResponse ComposerStage ClassifierAction SelectorEscalation ClassifierSession SummarizerVerifierRetrievalCustom
Versions and release control
Know what changed, and what is live.
Content saves carry an author and change note. Compare versions in split or unified view, stage selected versions in Dev or Staging, promote to Production, or restore an earlier version without deleting history.
Dev
v14
Scratch pointer
Staging
v13
Under review
Production
LIVEv12
Live for users
Release history records every promotion Production records each explicit release promotion
Line-level comparison
Compare any two versions with added, removed, and unchanged line counts.
Non-destructive rollback
Promote an older version back to production while preserving everything newer.
Runtime attribution
Traces record which prompt and version were used, so results map back to source.
Important release behavior
Saving content creates and promotes the new current version. Keep a prompt in Draft while authoring if changes must not reach runtime. Dev and Staging are review pointers; Production records the explicitly promoted release version.
Test and improve
Use real responses and outcomes, not prompt-engineering instinct.
Run a prompt directly in the playground, compare two versions on real traffic, or let the Pro optimization pipeline propose and score candidates from production and test samples.
Live playground
BYOK · real response · nothing saved
Resolve variables with test values, send a user message, and inspect the response, model, token usage, and final system prompt before using it with customers.
USER MESSAGE
Can I refund an order purchased 45 days ago?
MODEL RESPONSE
Your standard return window is 30 days. I can help check whether an exception applies…
Version A/B testing
Stable assignment · real traffic · optional promotion
Split traffic between two explicit versions. Identified users remain on the same arm, and the collector evaluates sample counts, metric lift, p-value, and confidence interval before selecting a winner.
A · control
50%
traffic allocation
B · candidate
50%
traffic allocation
Pro and higher
Five optimization strategies
01
OPRO
Uses scored history to ask the model for stronger prompt variants.
02
APE
Generates multiple candidates, evaluates them, and selects the best result.
03
DSPy Bootstrap
Builds optimized few-shot instructions from successful examples.
04
PromptBreeder
Mutates task prompts and improvement instructions across candidates.
05
Deterministic
Applies fast local prompt rules without making an optimizer model call.
Model-based optimization uses your configured provider key. Set the evaluation metric, minimum sample size, improvement threshold, candidate count, environment scope, and whether a winner may auto-promote.
Operate prompts as shared infrastructure
Visibility for builders. Boundaries for production.
Prompt quality is not only the text. Rylvo also shows who can edit, where a prompt is used, which traffic produced its metrics, what each model call cost, and how a release reached production.
Prompt-level editors
Owners and admins retain control while named workspace members can be granted edit and optimization access.
Reusable scope
Keep a prompt bot-specific, org-level, inherited, stage-specific, or shared across an agent group.
Dependency visibility
Inspect the bots, groups, parent prompts, and org-level relationships affected by a change.
Production / test lens
Keep playground and test traces separate from live success, latency, volume, and cost attribution.
Import
Create a fresh quota-checked prompt from a portable Rylvo JSON envelope.
Export
Download prompt source, metadata, placeholders, and version history for portability or review.
Prompt management FAQ
Know the runtime behavior before you publish.
Clear answers about saving, testing, deployment, experiments, and plan availability.
Does saving an edit immediately affect production?
Saving prompt content creates a new numbered version and makes that version current. If the prompt is Active and attached to a live bot, the updated content can be used by subsequent turns. Use Draft status while authoring, and use the version and environment tools for controlled promotion or rollback.
Can I test a prompt with a real model?
Yes. The live playground resolves your placeholder values, sends the prompt and a test user message to the selected model, and displays the response, token usage, and resolved system prompt. It uses your provider key and does not save the conversation.
How are prompts selected at runtime?
The engine loads Active prompts assigned to the bot plus eligible group-shared prompts. It resolves parent inheritance and placeholder defaults, applies workflow-stage scope, and records the prompt and version on the resulting trace.
Can I compare or restore an older version?
Yes. Select any two versions for a split or unified line-level diff. Promoting an older version makes it current again without deleting newer history.
How do prompt A/B tests work?
Choose two versions, a traffic split, evaluation metric, sample requirement, and improvement threshold. Identified users remain on a stable arm during the test. Results include sample counts and statistical evidence, with optional winner promotion.
Which plans include prompt optimization?
Core prompt management is available on every plan within each plan’s prompt limit. Self-improving prompt optimization is available on Pro and higher and uses your configured model-provider key.
Core prompt management on every plan
Make every instruction inspectable and reversible.
Start with a Draft, run it against your own model key, and move to an Active production version when the response is ready.