templates

Skills and advisors

The Roxels agent is not a single LLM with a fixed bag of tricks. It's a coordinated set: a speaker that talks to the user, skills the speaker can invoke as tools, and advisors that observe and improve the conversation in real time.

You attach skills and advisors to a template. Each comes with its own prompt section that's added to the agent's working brief.

Skills

A skill is a discrete capability — something the speaker can call when it's the right moment.

Examples of skills:

Skill	What it does
Send email	Drafts and sends an email mid-conversation.
Lookup customer	Queries your CRM by name/email to pull context.
Run walkthrough	Coordinates a screen-share walkthrough with step-by-step guidance.
Compose handoff	Generates a long, structured prompt to hand off to another agent.

Skills are typed and parameterized. The speaker LLM decides when to call them, with what arguments, based on the conversation. The result feeds back into the conversation and into goal extraction.

Attaching a skill

In the template's settings.speaker_skills array:

{
  "settings": {
    "speaker_skills": ["send-email", "lookup-customer"]
  }
}

Each skill name corresponds to a registered skill in your org. The dashboard's skill picker shows what's available.

Skill sets

A skill set is a curated bundle of skills, useful when several skills go together (e.g. a "customer support" bundle that includes lookup, send-email, and create-ticket). Attach a skill set with settings.speaker_skillsets.

Custom skills

You can create custom skills in the dashboard. A custom skill has:

A name and description — shown to the LLM.
A prompt section — added to the agent's brief when the skill is attached.
A tool schema — what arguments the skill accepts.
A handler — what the skill does (typically calls your API).

For org-specific capabilities, this is the path. See the dashboard's skill authoring guide.

Advisors

An advisor is an observer agent that watches the conversation and feeds the speaker agent with structured concerns — but never talks to the user directly.

Examples of advisors:

Advisor	What it watches
Missed extraction	Notices values the user said that weren't captured to a goal.
Goal stuck	Notices when the speaker has been on the same goal too long.
User confusion	Notices when the user keeps asking for clarification.

Advisors don't speak. They emit concerns — short structured notes the speaker sees on the next turn. The speaker decides whether to act.

Why advisors instead of just a longer system prompt?

Because the speaker's context is finite and expensive. Loading every conceivable rule into the speaker's brief slows down every turn. Advisors are separate, cheaper LLMs running in parallel — they see the conversation and only inject relevant guidance.

This is also why advisors don't talk directly: a chorus of voices would confuse the user. The speaker stays the single voice; advisors guide silently.

Attaching an advisor

In the template's settings.advisors array:

{
  "settings": {
    "advisors": ["missed-extraction-watcher", "user-confusion-watcher"]
  }
}

Speaker / observer model split

By default, the speaker and observer (advisor) agents run on tuned models — different speeds, different cost profiles. You generally don't pick the model; it's set at the org or system level for the right balance of speed and quality.

If your org has model overrides, the dashboard shows them under template settings (advanced). Don't change these unless you have a measurable need.

When to use what

Situation	Reach for
The agent needs to perform an action (call your API, send an email)	A skill
The agent keeps missing or mishandling something	An advisor
You want to ensure a behavior across many templates	A skill or advisor set, applied broadly
The behavior is one-template-specific and lightweight	A line in `settings.system_instructions`

Skills extend capability. Advisors extend judgment. System instructions extend voice. Most templates use all three in moderation.