Working with agents

Skills

Reusable instructions an agent loads when a request matches. Record one, or bring in one somebody else wrote.

A skill is a folder with a SKILL.md in it. When a request mentions one of the skill's trigger terms, its instructions are added to that turn, so the agent follows a procedure you defined instead of improvising one.

Skills live under ~/.pragma/skills/, one directory each, and are re-read every turn. A skill you add works on your next message, with no restart.

#What Pragma already knows

Every bot comes with Pragma's own know-how for the work people ask for most: launch films and promo videos, posters and social graphics, websites, 3D, and your phone. You do not install it or pick it: ask for the thing, and the bot uses what the job needs, explains how it will go, and says when a stronger model would do it better.

Some of that work needs tools on your computer. Videos need Node.js 22 or newer and FFmpeg, and 3D shots need Blender. When one is missing, the bot tells you what it is for and where to download it, and installs it only if you say yes.

#Bringing in one somebody else wrote

Open Teach a skill and switch to the Skills tab. Three ways in:

• Choose a folder. The whole skill, including any references/ images its instructions point at.
• **Paste a SKILL.md.** Instructions only, no attachments.
• Record one yourself. Below.

Skills written for other agents work here unchanged. Anthropic's Claude Skills, Google Antigravity folders and OpenAI agent folders all use the same shape: a SKILL.md opening with YAML frontmatter carrying at least a name and a description. Pragma reads that frontmatter and writes the manifest its loader needs.

#Trigger terms

The one thing no other format declares is when a skill should fire. Claude decides by reading descriptions. Pragma matches trigger terms against the text of a turn.

So an imported skill gets terms derived from its name and description, and the confirmation tells you which ones were chosen. If a skill never seems to load, that list is the first thing to check. A skill that already carries its own trigger terms keeps them.

#Seeing what is installed

The same tab lists the skills you added, with their trigger terms, and removes them.

#Recording one

Show the app how to do something once and it writes the skill for you. Speak while you work: your narration becomes the body of the instructions.

What gets captured depends on the platform, and the page says so before you start:

The per-action detail on macOS comes from a signed helper using the Accessibility API, which is also what lets it name the button you clicked and suppress anything typed into a password field. There is no equivalent on Windows without new native code, so a recording there is screen frames every few seconds plus what you said.

That makes narration the thing that carries a Windows recording. Say what you are doing and why, not just what you are clicking.

#Turning your narration into text

The page picks one of three from what is set up. There is no setting to choose between them:

• An AssemblyAI key, saved under Settings, then Connections, transcribes live while you record.
• On this computer (Windows and macOS): Whisper runs on your own computer after you press Stop. Set it up once with Set up transcription on the recorder, or under Settings, then Connections, then On this computer. That downloads a speech model from huggingface.co into ~/.pragma/models/whisper: 190 MB for the more accurate model, or 60 MB for the faster one. After that it works offline. English is the default; Other languages switches to the multilingual model, which detects the language. You can record while it downloads.
• Neither: your voice is still saved with the skill, and you type what you did and why.

On this computer, the review screen opens as soon as you stop, and the text fills the Narration box when Whisper finishes. A ten-minute narration usually takes from under a minute to a few minutes, depending on the computer, and Skip stops waiting.

Whichever way the text arrives, you can fix names it got wrong before creating the skill. It is the first thing an agent reads.

#Privacy

Raw keystrokes and clipboard contents are never stored. Screen frames can contain whatever was on screen and stay on this computer. Your voice stays on this computer too, including when it is transcribed on this computer: the speech model download is the only network request, and it carries no audio. Only with an AssemblyAI key saved is microphone audio streamed to AssemblyAI, for the live transcript, and nowhere else. Everything is shown to you for review before it becomes a skill.