RylvoRylvo

Skills

Reuse the behavior. Load it only when needed.

Turn agent instructions into scoped, testable Skills. Preview exactly what will activate, inject only relevant guidance per turn, trace every match, and reverse any bad change.

50

editable catalog skills

15

catalog categories

4

lifecycle states

1

runtime skill type today

Skill workflow

Package the behavior. Load it only when relevant.

Rylvo Skills turns reusable agent instructions into managed runtime units: scoped, previewable, selectively activated, traced, versioned, and reversible.

Per-turn activation

message → score → select → inject

Skills layer on

User message

“I want a refund and my money back. Cancel order. I am angry and frustrated; this is ridiculous.”

The same message can activate more than one relevant skill.

Match candidates

min 0.25
Refund policy0.67
De-escalation0.39
Order tracking0.06
system.prompt

## ACTIVATED SKILLS

Refund policy

Ask for the order number and check the return window…

De-escalation

Acknowledge frustration and offer one clear next step…

Only Active skills are eligible
Highest scores are kept first
Selection is recorded on the trace
01

Create

Write a reusable instruction package.

Start manually, install an editable catalog copy, import JSON, ask Architect to create one, or optionally promote a cluster of related agent rules into a Draft skill.

02

Scope

Choose exactly which bots can use it.

Keep the skill bot-specific, make one org-level skill available to every bot, or clone an existing skill into another bot as an independently editable copy.

03

Preview

Test activation before live traffic.

Enter a sample message and see which skills match, their score, matched keywords, scope, status, instructions, and why they were selected.

04

Activate

Only Active skills enter the runtime.

Enable the Skills Layer for the bot. Matching Active prompt-template skills are ranked, capped, and added to the system prompt for that turn under ACTIVATED SKILLS.

05

Observe

See what fired and how the turn ended.

Traces carry the skill ID, name, match route, and score. Usage events roll up activations, operational success and failure, confidence, and last-used time.

06

Control

Change safely and reverse regressions.

Every edit creates a version. Compare the recorded history, restore an earlier version, disable or archive a skill, and optionally auto-disable high-failure skills.

Inside a skill

Instructions, activation, scope, and lineage in one unit.

Author the behavior and when it applies, then keep its scope, status, source, versions, expected tools, and operational counters attached to the same managed object.

Skills / Refund policyActive · v4
System instructionsPrompt template

# Refund handling

1. Ask for the order or transaction ID.

2. Check the configured return window.

3. Explain eligibility and next steps.

4. Never promise an exception.

Required tools are expectations, not an execution guarantee.

refundmoney backreturn policycancel order
Scope

All bots

Status

Active

Version

v4

Source

Prebuilt fork

Instructions

The prompt guidance injected when the skill activates.

Triggers

Natural-language conditions plus automatically extracted keywords.

Scope

One bot, all bots, or an independently cloned bot copy.

Operational counters

Activations, success/failure, confidence, and last use.

Runtime boundary

Prompt Template is the only skill type that can be created and activated at runtime today. Tool Chain, Code Snippet, and Workflow remain schema types marked coming soon. Required tools and tool-chain steps on a prompt-template skill are advisory hints shown to the model; Rylvo does not execute them as a deterministic chain.

Matching and preview

Tune activation without guessing what will fire.

Preview uses the canonical lexical scorer, while live traffic adds optional Lite+ semantic similarity. Configure the balance, threshold, and per-message cap for each bot.

Match preview

same lexical scorer as runtime

Test message

I want a refund and my money back. Cancel order. I am angry and frustrated; this is ridiculous.

67%

Refund policy

keyword: refund

39%

De-escalation

condition: angry

Matching and activation

Keyword weight

Literal keyword coverage

0.40

Trigger weight

Natural-language trigger hits

0.60

Minimum score

Required to activate

0.25

Max per message

Highest-ranked skills kept

8

Semantic matching · Lite+

Optional embedding similarity adds paraphrase coverage. If embedding generation fails, the engine falls back to lexical matching.

Always-on option

An Active skill with instructions and no triggers activates every turn.

Score-ranked cap

The highest-scoring matches are kept up to the configured maximum.

Trace attribution

Each activated skill carries its ID, name, match reason, and score into the trace.

Prebuilt library

50 working starting points—not locked examples.

Browse 15 categories, inspect the complete instructions and triggers, then install a Draft or Active fork at org or bot scope. Every copy remains editable and quota-checked.

Skill Library

50 templates15 categoriesSearch · filter · install

Compliance

3

Content & Marketing

3

Conversation Design

3

Customer Support

5

Data Collection

3

E-commerce

4

Finance & Insurance

3

HR & Internal

3

Healthcare

4

IT & Technical

3

Knowledge & Escalation

2

Multi-Agent

5

Real Estate & Education

2

Sales

4

Utilities

3

You own the installed copy

Edit it without changing the catalog source.

Install by scope

Choose org-level, one bot, or several independent bot copies.

Quota checked

Every installed copy counts toward the plan’s skill limit.

Lifecycle and governance

Operate skills like production configuration.

Keep inactive work out of runtime, preserve every change, move skills between workspaces, and use runtime health signals to catch harmful behavior early.

Lifecycle

01

Draft

Author and preview

02

Active

Eligible to match

03

Disabled

Paused, reversible

04

Archived

Retained, inactive

Version history

Updates capture instructions, triggers, status, reason, actor, and counters at change time.

Non-destructive rollback

Restoring an older snapshot creates a new version instead of erasing newer history.

JSON portability

Export all skills or a bot’s effective set; import up to 50 items with per-item validation.

Clone to another bot

Create a separate bot-scoped copy with fresh counters and its own version history.

Safety at write and runtime

Create, update, import, and catalog-install paths reject known prompt-injection patterns. The runtime scans instructions again before rendering and records suspicious content in logs.

Operational telemetry

Idempotent per-turn events are asynchronously aggregated into activation, operational success/failure, average confidence, and last-used counters without putting a hot write on the skill document for every request.

Auto-disable protection

Configure an org-wide failure threshold and minimum usage window. The scheduled guard can disable an Active skill that crosses the threshold, append a version record, and notify admins.

Export JSON

Import JSON

Free

3 skills

Lexical matching

Lite

25 skills

Lexical + optional semantic

Pro and higher

Unlimited skills

Lexical + optional semantic

Skills FAQ

Know exactly what activates—and what does not.

Clear answers about runtime support, matching, metrics, always-on behavior, and plan availability.

What does a Rylvo Skill do today?

A live Rylvo Skill is a prompt-template capability: instructions plus activation triggers, scope, lifecycle, versions, and runtime telemetry. When an Active skill matches a message, its instructions are added to the bot’s system prompt for that turn.

Do tool-chain skills execute tools automatically?

No. Prompt-template skills are the only runtime-supported skill type today. A skill may name required tools or include a suggested tool sequence, but those are instructions for the model—not a deterministic tool-chain executor. Code-snippet and workflow types are also schema-only and marked coming soon in the dashboard.

How does matching work?

The default lexical score combines keyword coverage and trigger-condition hits. The defaults are 40% keyword and 60% trigger weight, a 0.25 activation threshold, and up to eight activated skills. Operators can tune all four values. Lite and higher can optionally add embedding similarity so paraphrases can match.

Can a skill run on every turn?

Yes. An Active skill with instructions and no trigger keywords or conditions is treated as always-on by the live runtime. Triggered skills remain dormant until they clear the configured score.

What do the success and failure counters mean?

They are operational turn outcomes—not a custom business KPI. A normal non-blocked turn counts as success; engine errors or blocked turns count as failure. Confidence also accounts for the match score and guardrail interventions.

Which plans include Skills?

Skills are available on every plan within quota: Free includes 3 skills, Lite includes 25, and Pro and higher include unlimited skills. Semantic matching is available on Lite and higher; Free uses lexical matching.

Skills on every plan

Stop copying behavior into every bot.

Install a Draft from the catalog or author one capability, preview the exact match, activate it deliberately, and keep every later change reversible.

Start free Compare plans