Skills

A skill is your company’s way of working, written for an agent: the steps of a task, the rules to respect and the format of the result. It can carry supporting files, such as examples, tables or templates. You write it once and connect it to every agent that needs it, instead of repeating it in each agent’s prompt.

How the agent uses it

The agent does not read all its skills in full at every step. When it starts, it only sees the name and description of each one. When a job matches a description, it opens that skill, reads it in full and follows it. Thanks to that you can connect several skills to one agent without making it slower or more expensive.

That is why the description is the most important part of a skill: if it does not say what the skill is for, the agent will not know when to open it.

Where they are created

• Your skills, in My Skills (under Configuration). You can write the instruction on the page itself or upload a ready-made skill as a .zip or .skill. See My Skills. In the flow they are connected with Stored Skill.
• NappAI skills, prepared for common services and tasks: Slack, WhatsApp and Telegram, Google Workspace, Microsoft 365, HubSpot, Jira, e-commerce, data analysis, research and others. They are in the Skills family of the palette (ManagedSkill).

In both cases, the component’s Skills output connects to the Skills input of the Deep Agent, which appears when you turn on Enable Skills.

On a voice agent

The same skills work on a Voice Agent: turn on Enable Skills in its Advanced tab and connect them to the Skills connector that appears. The agent works as the Deep Agent does: it sees the names and descriptions and opens a skill when what the caller asks matches it. There are three differences:

• It only reads the skill’s text. No script runs on a call.
• What it reads stays with it for the rest of the call, so it does not open it again.
• The first time it uses a skill in a call, that turn takes a little longer to answer, usually under a second.

For voice, keep the skill short and to the point: the steps, the rules and the facts the agent has to say. The Voice Agent Realtime takes no skills.

How to write a good skill

• A name and a description that say when to use it. “Our method to suggest the name of a new business and check its domains” works; “naming” does not.
• Numbered steps, in order. The agent turns them into its work plan.
• Concrete rules you can check. “Two joined words, easy to say in English and in Spanish” is either met or not; “a nice name” cannot be checked.
• What to do when something goes wrong. “If both domains are taken, replace that name with another one and check it too.” Without that sentence, the agent delivers whatever it found even if it is no use.
• What to do when a tool fails. “If a check fails, write ‘not checked’: never call it free or taken.” A model tends to fill the gap with the most likely answer; a rule like this makes it say only what it really knows.
• The format of the delivery. Sections, table columns, length. That way every report comes out the same.
• Tools by what they do, not by their technical name. “Search the internet” and “check the domains” still work if you change search engine tomorrow.
• One task per skill. If a skill starts explaining two different jobs, split it in two.

A skill looks like this (it is the one in the example):

# Naming study

When you are asked for the name of a new business, follow these steps in this order:

1. Search the internet for three similar businesses in the same city, so you neither repeat nor imitate their names.
2. Suggest five names made of two joined words, easy to say in English and in Spanish, without the name of the city.
3. Check the .com and .net domains of each name, written in lower case, with no spaces. If both are taken, replace that name with another one and check it too, until you have five names with a free domain. If a check fails, write "not checked": never call it free or taken.
4. Deliver the report in this format:
   - **Competitors**: the three businesses you found, one line each.
   - **Names**: a table with the columns Name, .com, .net and Why it works.
   - **Recommendation**: the name with the most free domains and one sentence on why.

Skills and prompt: what goes where

In the agent’s System Prompt	In a skill
Who the agent is and who it works for	How one specific task is done
Rules for everything (language, never make up data)	Steps, rules and format of that task
”Before starting, check whether a skill explains the task and follow it”	Whatever is repeated across agents

A short prompt and several well-described skills make agents that are easier to maintain: to change how a task is done, you change the skill, and every agent that uses it applies it on its next run.