Aevum Cadence AI

Your first skill

Skills are how your assistant learns to do new things. Not a prompt you type once and forget: a durable recipe it can follow again.

Canonical URL: /docs/getting-started/your-first-skill

Why skills matter

Most people use AI as a chat. Ask, get an answer, move on. The next day the same question starts from zero. That is architecture, not model quality.

A skill is the opposite. You teach the workflow once (format, voice, tools, edge cases). It runs the same way next week.

How skills sit next to tools: Skills and tools. Platform feature areas: Features.

What's a skill?

A skill is a folder. At minimum it has a SKILL.md (purpose, when to use it, the steps). It can also carry scripts, reference notes, and assets.

You do not have to think about skills on every turn. The assistant picks one when your request matches. Knowing they exist, and that you can write your own, changes what is possible.

Built-in skills

The desktop app ships a first-party catalog (communication, productivity, scheduling, development, and more). Some work with no setup. Others need a one-time connection (mail, chat, phone), and the assistant should walk you through it on first use.

Do not treat a remembered list as the catalog. Open Skills reference for the pages that exist, or ask the running assistant to list installed skills.

Installing a skill that is not built in

Portable skills follow the Agent Skills layout. You can drop a skill folder into the assistant's skills directory, or ask the assistant to find and install one with your permission.

"Search for a skill that drafts weekly status notes."

Read what it does before you allow install. A skill is code plus instructions running with the assistant's tools.

Building a custom skill

Describe the job in chat. Treat it as teaching, not a feature request. The assistant can draft the recipe (Skill Management) and save it.

Step 1: Pick one task

Pick something you do every week and find tedious. Specific, not generic. Not "write blog posts." Something like:

"Check competitor pricing every Monday and give me a table with price, change, and verdict."
"Read overnight support tickets, group them by category, and post a summary at 9 AM."

Narrower is better. One job, one output.

Step 2: Draft the recipe

A good skill has four ingredients:

  1. Clear purpose. One task, one output.
  2. Workflow steps. The sequence every time.
  3. Tool access. What it may read, search, or call.
  4. Output format. What a perfect result looks like. Show an example.

Be specific about format, voice, and edge cases. Secrets still go through the credential prompt, never into SKILL.md.

Step 3: Test and iterate

Run it. The first pass is rarely perfect. Tell the assistant what is off (missing column, wrong tone, a name you forgot) and have it patch the files. Three rounds usually gets you to reliable.

Step 4: Refine until reliable

Custom skills live as plain text under the instance workspace skills/ directory (see The workspace). You can:

A reliable skill is one you trust enough to run without checking the output first.

If a custom skill does not load, ask why. Invalid SKILL.md is the usual cause. See Common issues.