Rylvo Agent Skills: Hermes-Style Reusable Capability Packages for Self-Evolving AI Bots

Most AI bots are built as monolithic prompt blocks. A single system prompt contains the bot's identity, personality, constraints, and every procedure it might need. This works for simple bots, but as use cases grow, the prompt becomes bloated, expensive, and ineffective. The LLM spends tokens on instructions it does not need, gets confused by conflicting guidance, and cannot adapt to different conversation contexts.

Rules help — they add constraints like "always confirm order numbers before refunds." But rules are reactive and flat. They tell the bot what not to do, not how to do something well. They are always injected, even when irrelevant, and they carry no awareness of the tools or workflows the bot should use.

Skills are different. A skill is a reusable capability package that bundles system instructions, trigger conditions, required tools, tool chains, and effectiveness metrics into a single coherent unit. When a user asks about refunds, the refund handling skill activates automatically. When they ask about bookings, the appointment scheduling skill activates instead. Only relevant skills are injected, keeping the context tight and the responses precise.

Rylvo Agent Skills is a Hermes-style capability system built into the Rylvo platform. It supports four skill types, five creation sources, a full lifecycle from draft to archive, lexical matching on a sub-5-millisecond hot path, auto-promotion from rule clusters, immutable version history with rollback, per-skill effectiveness tracking, and both bot-specific and organization-level scoping — all managed through the Skills dashboard at /dashboard/skills.

In this guide, we will explore the complete Skills system: what skills are, how they differ from rules, the four skill types, the matching algorithm, runtime injection, the dashboard, auto-promotion, versioning, effectiveness tracking, and the self-evolution loop that turns reviewer insights into new capabilities.

What Are Skills? Capability Packages, Not Constraints

A skill is a complete capability that tells the bot both what to do and how to do it. Unlike a rule, which is a single text constraint, a skill bundles multiple components into a coherent package.

System Instructions are the full procedure injected into the system prompt when the skill matches. They are not one-line constraints — they are step-by-step guidance. A refund skill might say: "When the user asks for a refund, confirm the order number first, check the return policy window, then offer store credit or full refund based on eligibility."

Trigger Conditions are natural-language sentences that determine when the skill should activate. "refund request," "return policy," "money back" — these phrases tell the matching engine when this skill is relevant.

Required Tools declare which connector or MCP tools the skill expects the bot to use. A refund skill might require refund_create and order_lookup tools, nudging the LLM toward correct tool selection.

Tool Chain is an ordered sequence of tool calls with parameter mapping, rendered as guidance in the injected prompt.

Effectiveness Metrics track how often the skill is used, how often it succeeds, and the average confidence of matches.

Skills are org-scoped but can be bound to a specific bot or shared across the entire organization. Bot-specific skills only activate for their assigned bot. Org-level skills activate for every bot in the organization — ideal for brand voice, compliance procedures, and common support workflows.

Four Skill Types for Different Capabilities

Rylvo Skills supports four distinct skill types, each designed for a different class of capability.

Prompt Template

The most common skill type. A prompt template skill injects a full procedure into the system prompt when matched. It is ideal for domain expertise, support workflows, and procedural knowledge. When a user asks about returns, the returns skill activates and the bot knows exactly what steps to follow. When they ask about something else, the skill stays dormant and does not clutter the context.

Tool Chain

A tool chain skill declares an ordered sequence of tool calls the bot should execute. It maps parameters between steps and guides the LLM through multi-tool workflows. For example, a booking skill might specify: first call calendar_availability, then call appointment_create with the selected slot. The tool chain is rendered as structured guidance in the injected prompt.

Code Snippet

A code snippet skill encapsulates programmatic logic that the bot can reference. This is ideal for calculations, data transformations, or conditional logic that would be verbose to describe in natural language. The code is rendered in the prompt as reference material, guiding the LLM's reasoning.

Workflow

A workflow skill captures multi-step procedures with branching logic. Unlike a simple prompt template, a workflow skill can describe decision trees: if condition A, do X; if condition B, do Y. This is ideal for complex business processes that require conditional handling.

Five Sources: From Manual Creation to Auto-Promotion

Skills can originate from five different sources, giving organizations flexibility in how they build their capability library.

Manual skills are created by human operators through the Skills dashboard. This is the most common path — an operator identifies a recurring need, writes the instructions, defines trigger conditions, and activates the skill.

Generated from Rules skills are auto-promoted by the reviewer engine. When the Agent Evolution system detects a cluster of related active rules, it bundles them into a skill draft. This bridges the gap between low-level rules and high-level capabilities, turning scattered constraints into coherent procedures.

Imported skills come from another bot, organization, or a future Skills Hub. This enables reuse across teams and sharing of proven capabilities.

Architect skills are created by the Workspace Architect during bot setup interviews. When an operator describes their business process, the Architect can generate matching skills automatically.

Reviewer Suggested skills are proposed by one of the four reviewer agents during the evolution review process, identifying gaps that would benefit from a structured capability package.

The Matching Algorithm: Sub-5-Millisecond Hot Path

Skills do not help if they are slow. Rylvo uses a lightweight lexical matching algorithm that runs in under 5 milliseconds, avoiding expensive LLM or embedding calls on the hot path.

For each active skill, the engine extracts up to 30 trigger keywords from the skill's trigger conditions. When a user sends a message, the engine scores every active skill using two signals: keyword overlap counts how many trigger keywords appear in the user message, and trigger phrase overlap splits trigger conditions into phrases of 4 or more characters and counts substring hits. The final score is a weighted combination: 40% keyword score plus 60% trigger score.

Only skills with a score of 0.25 or higher are considered candidates. The top 5 matches are returned, sorted by score, and injected into the system prompt. Skills with no match are completely ignored, keeping the context window tight and costs low.

This lexical approach is intentionally fast. A future semantic matching option using embeddings is planned for cases where paraphrased queries share no keywords with trigger conditions, but the current system prioritizes speed and reliability.

Runtime Injection: Where Skills Fit in the Prompt

On every bot turn, the runtime augments the system prompt in a specific order. Skills are injected after knowledge base retrieval and rule injection, but before episodic memory recall and user dialectic models. This order ensures that grounded facts and constraints are established first, then skills add procedural guidance, then personal context and memory complete the picture.

Matched skills are appended as a structured ## LEARNED SKILLS (Auto-Matched — HIGH PRIORITY) section. Each matched skill shows its name, type, instructions, and expected tools. This explicit formatting tells the LLM that these are high-priority, contextually relevant procedures that should guide its response.

Only matched skills are injected. Unlike rules, which are always present if active, skills are conditionally injected based on the conversation context. This keeps the system prompt shorter and more focused than a dump of all possible rules.

The Skills Dashboard: Complete Capability Management

The Skills dashboard at /dashboard/skills provides full CRUD and lifecycle management for all skills in your organization.

Bot Selector and Grouped View

A bot selector lets you filter skills by specific bot or view "All Bots." When viewing all bots, skills are grouped by bot with expandable sections. Org-level skills appear in their own "Organization-wide" section at the top. Each group shows the bot name, avatar color, and skill count.

Status Filters and Search

Status filter pills let you view All, Active, Draft, Disabled, or Archived skills. A full-text search bar searches across skill names, descriptions, system instructions, trigger conditions, and tags — making it easy to find the right capability even in large libraries.

Expandable Skill Cards

Each skill appears as a card showing its name, type badge, status badge, bot assignment, description, and usage statistics. Click any card to expand and see the full system instructions in a monospace scrollable block, trigger conditions as badge-style tags, required tools as tag chips, auto-promotion lineage if applicable, version number, source attribution, and usage count with success rate.

Lifecycle Actions

Every skill supports a full lifecycle. Draft skills are created but not injected. Active skills enter the matching pool and can be injected on future turns. Disabled skills are temporarily paused and can be re-activated instantly. Archived skills are retired but preserved for audit. Skills can also be deleted permanently with confirmation.

Create Skill Modal: Building New Capabilities

Creating a skill manually is straightforward. The modal captures: name and description, skill type selector, system instructions editor, trigger conditions (one per line), required tools (comma-separated), and an org-level checkbox to share the skill across all bots.

Skills are created in draft status by default. Operators review the skill, then activate it when ready. This prevents untested capabilities from entering production automatically.

Auto-Promotion from Rules: The Self-Evolution Bridge

One of the most powerful features of Rylvo Skills is the bridge between Agent Evolution and runtime capabilities. When auto-skill promotion is enabled in Agent Evolution Settings, the reviewer engine analyzes clusters of active rules.

The process works as follows. The engine clusters active rules by target and Jaccard text overlap with a threshold of 0.25. When a cluster reaches the configured promotion size (default 3 rules), the system proposes a skill draft. The draft combines the related rules into a single coherent capability package with system instructions, trigger conditions, and required tools. An operator must manually review and activate the draft before it enters production.

This bridges the gap between low-level constraints and high-level capabilities. Instead of maintaining dozens of overlapping rules, operators get focused, reusable skills that capture entire procedures.

Version History and Rollback

Every update to a skill creates an immutable version snapshot in a subcollection. Version types include created (initial snapshot), instructions updated, trigger conditions updated, and status changed. The full history is preserved for audit and comparison.

If a skill's performance degrades after an update, operators can roll back to any prior version with a single click. Rollback restores the skill to the selected version and resets effectiveness counters, so the rolled-back version starts with clean metrics. This gives operators the confidence to iterate on skills without fear of breaking production behavior.

Effectiveness Tracking: Data-Driven Skill Management

After each turn where a skill was matched, the system records usage metrics: usage count (how many times the skill was matched), success count (turns that completed without guardrail blocks or errors), failure count (turns blocked by guardrails or encountering errors), average confidence (running average of match confidence), and last used timestamp.

These metrics feed into operator decisions. If a skill has high usage but low success rate, its instructions may need refinement. If a skill is never matched, its trigger conditions may be too narrow. If a skill succeeds consistently, it is a proven capability worth sharing across bots.

Skills vs. Rules: The Key Differences

Rules and skills serve different purposes in the bot augmentation stack. Rules are single text constraints that always inject if active. They tell the bot what not to do. Skills are full capability packages that only inject when matched. They tell the bot how to do something well.

Rules have no tool awareness, no version history, and no effectiveness tracking. Skills declare expected tools, maintain immutable version history with rollback, and track usage and success metrics. Rules are promoted from reviewer insights one at a time. Skills are promoted from rule clusters, representing higher-level capability bundling.

Both systems work together. Rules establish baseline constraints. Skills add contextual capabilities. The combination gives bots both safety and competence.

The Self-Evolution Loop

Skills close the loop between observation and capability. Here is how it works end to end.

A bot runs conversations and generates traces. The reviewer engine analyzes recent runs, detects patterns, and generates rules. Over time, related rules cluster together. When a cluster reaches the promotion threshold, the system creates a skill draft. An operator reviews and activates the skill. On future turns, the skill matches when relevant and injects its procedure. The bot's responses improve. Usage metrics inform further refinement. The cycle repeats.

This is not just automation — it is a human-in-the-loop capability factory where AI identifies patterns, humans approve capabilities, and the runtime adapts automatically.

Enabling the Skills Layer

Skills do nothing unless the feature is enabled. Per bot, via Agent Evolution Settings, three toggles control the Skills layer: skillsEnabled is the master toggle that enables skill matching and injection, autoSkillPromotion enables the reviewer engine to propose skill drafts from rule clusters, and skillPromotionClusterSize sets the minimum number of rules in a cluster before promotion (default 3).

These settings are exposed in the Agent Evolution dashboard under the Hermes-Style Self-Improvement settings tab. Operators can enable skills gradually: first manual creation, then auto-promotion once they are comfortable with the system.

Performance: Caching and Dual-Path Execution

Active skills are cached in memory for 30 seconds keyed by organization and bot ID. This prevents Firestore round-trips on every message in a busy conversation while still reflecting recent activations within a reasonable window.

The skill engine runs on both client-side (for dashboard chat) and server-side (for API consumers). Both paths share the same matching logic and injection format, ensuring consistent behavior regardless of how users reach the bot.

Comparison: Skills vs. Static Prompt Engineering

Getting Started

Step 1: Enable Skills

Navigate to /dashboard/agent-evolution and open the Settings tab. Enable the Skills Layer toggle under Hermes-Style Self-Improvement.

Step 2: Open the Skills Dashboard

Navigate to /dashboard/skills. Select a bot from the dropdown or view all bots.

Step 3: Create Your First Skill

Click New Skill. Choose Prompt Template as the type. Name it (e.g., "Refund Handling"), write the system instructions (step-by-step procedure), add trigger conditions (one per line), and declare required tools. Save as draft.

Step 4: Activate the Skill

Review the skill in the dashboard. When ready, click Activate. The skill enters the matching pool and will be injected on future turns where triggers match.

Step 5: Monitor Effectiveness

After the bot has handled some conversations, check the skill's usage count and success rate. If the success rate is low, edit the instructions or trigger conditions and create a new version.

Step 6: Enable Auto-Promotion (Optional)

Return to Agent Evolution Settings. Enable Auto-promote Skills from Rules. The reviewer engine will now propose skill drafts when it detects rule clusters. Review and activate promising drafts.

FAQ

What is Rylvo Agent Skills? A Hermes-style capability system that lets you create reusable skill packages — system instructions, trigger conditions, tools, and metrics — that activate automatically when matched to user messages.

How do skills differ from rules? Rules are single text constraints always injected if active. Skills are full capability packages only injected when contextually matched, with tool awareness, version history, and effectiveness tracking.

What skill types are available? Four: Prompt Template (injected procedure), Tool Chain (ordered tool calls), Code Snippet (programmatic logic), and Workflow (multi-step procedure with branching).

How does skill matching work? Lightweight lexical overlap scoring: keyword matches (40% weight) plus trigger phrase matches (60% weight). Skills scoring 0.25 or higher are candidates; top 5 are injected. Runs in under 5 milliseconds.

Where do skills appear in the prompt? After knowledge base retrieval and rules, but before episodic memory and user context, as a structured ## LEARNED SKILLS section.

Can skills be shared across bots? Yes. Org-level skills (botId = null) activate for every bot in the organization. Bot-specific skills only activate for their assigned bot.

What is auto-promotion? When enabled, the reviewer engine clusters related active rules and proposes skill drafts. Operators must manually activate drafts before they enter production.

Is there version history? Yes. Every update creates an immutable version snapshot. You can roll back to any prior version, which resets effectiveness counters.

How is effectiveness tracked? Per-skill metrics: usage count, success count, failure count, average confidence, and last used timestamp.

How do I enable skills? Go to Agent Evolution → Settings → Hermes-Style Self-Improvement and enable the Skills Layer toggle. Optionally enable auto-promotion.

Ready to Give Your Bots Real Capabilities?

Rylvo Agent Skills transforms bot behavior from static prompts into dynamic, context-aware capabilities. Instead of stuffing every possible procedure into a bloated system prompt, you create focused skills that activate only when needed. The matching engine keeps the hot path under 5 milliseconds. The dashboard gives you full lifecycle control. Auto-promotion turns reviewer insights into draft capabilities. Version history lets you iterate safely. Effectiveness tracking tells you what works.

Build a skill library that grows with your bot. Start with manual skills. Enable auto-promotion when ready. Watch your bot get better at handling refunds, bookings, troubleshooting, onboarding, and every other procedure your users need — one matched skill at a time.

Open Skills and create your first capability today.

Teach once. Match automatically. Improve continuously.