Guide
English
Search the docs⌘K
12.2

Skills

A skill is a preset someone already wrote down: how to break one line into shots, how to keep a character looking the same across shots, how to pace 15 seconds versus 60. DOTV keeps them in the skill library: ones other people published, and ones you wrote yourself. Normally the Agent only knows that these presets exist. Pick one and that whole preset is read in for that turn, and then it works to it — if you're new to AI video, borrowing a proven approach beats working it out from scratch. This section covers what is in there, picking one, writing your own, starring, editing and deleting, and what happens if you don't pick one.

What is in there

The puzzle piece in the composerClick it to open this list. Typing / in the composer opens the same one — with a search box on top that matches names, slugs and descriptions.
1 / 2
The puzzle button in the input area opens this list.

The row above the list is the filters:

  • The three tabs switch which pile you are looking at — Common (everything published, plus your own space's), Favorites (whatever you starred), Mine (the ones you wrote).
  • The dropdown at the top right filters by what the preset produces: image, video, audio or text — that is a field filled in when the skill is written.
  • The search box matches three things: the name, that /name tag, and the one-line summary. When the library holds a lot, the list comes a page at a time and the last row is Load more.

This page does not list what is in there. The library is data — anyone can write a skill and publish it — so any list printed here would go stale. The column you see when you open the panel is the current one, and the line under each row is what it is for.

Picking one

Two ways in, same result:

  • click the puzzle button in the input area, labelled Skill, and pick from the list;
  • type / in the input and the same list pops up on the spot, then keep typing to search.

Click a row (the whole row is clickable) and a chip with the preset's name appears in your message. It takes effect with that message; it is not a toggle. The Agent remembers the method in the following turns, so you do not have to attach it to every line.

Writing one yourself

The routine you only settled on around the tenth message — character sheet first, then these camera moves per shot, five seconds each — can be saved as a skill on the spot and called back with one word next time.

The Create button on the panel is a dropdown with two routes:

This one What happens
Create a new skill No page change: the Agent puts a card in the conversation asking what you want — work out a new skill from scratch (it questions you first, pinning down the subject, the register, the aspect ratio), turn this canvas into a skill (it reads the canvas and draws the method out of those nodes), update one of my skills, test one of my skills (it runs a frame or two from the skill's written instructions and tells you which lines were too vague to follow). Pick one and it writes and saves the thing itself — no form to fill in
Import an existing skill Reads in a .md file or a whole folder (the submenu picks file or folder) and opens that form with the content in it — the title comes from the file information or the first heading in the body, and is left blank rather than guessed

Import opens a new page (writing a skill is slow work, and this way not a word of the conversation on the canvas is disturbed — finish, switch back, carry on); Create a new skill stays in the conversation and never takes you off the canvas.

What the Agent saves is always private — only your own space sees it. There is no publish switch in the interface today, so nobody else finds these in the library.

To write a skill from nothing, neither of those two is the way: go to Skills in the left rail, where the Create button at the right of the tab bar opens an empty form.

The form asks for: Skill name, One-line summary, Skill content (the body the Agent actually reads; required), When to use it / How to use it / What it produces, a Type, and two optional boxes side by side at the end: Cover image and Sample clip (a clip that plays when someone opens this skill in the library). Save and it lands in the skill library — reachable from Skills in the left column, and on the Mine tab of the panel.

  • That one-line summary matters more than it looks: on an ordinary turn it is all the Agent reads of your skill, and it decides from that line whether this is the turn to use it. "Cut a voiceover script into a 15-second vertical ad" beats "my advertising tricks" by a mile.
  • You do not invent the /name tag: it is derived from the title. A title with no Latin letters in it derives nothing, so it falls back to something like skill-xxxx — usable, just harder to type from memory.

The Skill content field holds a small editor: Code / Preview to switch between the two, Expand for full screen, Upload for one file or a whole folder (an uploaded .md can replace the body outright), and Improve the skill content, which has the model rewrite what you wrote — it overwrites your text, so undo it if you do not like the result. In the file tree on the left, the skill body is pinned at the top and cannot be deleted: it is the core of the skill.

Starring, editing, deleting

Hover a row and a few buttons appear at its right end:

  • The star favorites it. For a skill someone else published, starring is the only way to keep it within reach.
  • Only the rows you wrote get two more: the pencil (opens that one for editing in a new window) and delete (it asks first). Skills other people published can only be starred and used.

Favoriting is more than a bookmark: the list the Agent sees on every turn holds only the ones you wrote and the ones you starred the most-used entries from each. So starring is the way to say "I want it reaching for this one from now on" — dumping the whole library on it would only crowd out the presets you actually use. Write nothing and star nothing and that list is empty, and the Agent has no preset to go on that turn.

What if you do not pick one

It still works. Every preset on that list has its name and one-line summary in the Agent's view on every turn, and it decides for itself whether to follow one — ask for something that lands squarely on one of those summaries and it will most likely reach for that preset.

The difference is how much of it the Agent reads: without picking, only that one line; pick it and the whole body comes in (shot list templates, rhythm patterns, the usual ways it goes wrong). So if you want it to follow a preset strictly, pick that preset once yourself.

Try it in DOTVReading only goes so far: open DOTV and make an AI video with the feature this section covers.