Publit.Si
Docs›Studio

Using Publit.Si

Studio: make Shorts with your agent

In the Studio, your coding agent works in your project's own repository, like in your terminal: with its render scripts, its CLAUDE.md and its skills. It makes the video, writes the texts and hands it over. Nothing goes online until you approve it.

You need

  • Claude Code, Codex, Cursor or OpenCode installed and signed in (see AI agents)
  • A repository with your render scripts, or an empty folder for a new one

Set the project folder

Don't want the agent near your app's code? Create a marketing folder makes a new folder just for the videos, with brand/, screens/ and music/, and links it. The agent stays inside it.

To switch later, use Change project folder in the chat's ⋯ menu.

  1. Open Studio and pick the project.

  2. Under Project folder (repository), search your repositories, use Choose in Finder… or type the absolute path. You do this once per project.

  3. Pick the repository that has your render scripts. Not the publitsi folder itself, and not your whole home folder.

Make a Short

An empty chat asks “What should we make today?” and offers quick actions, e.g. Make a new Short, Make a series, Short from a backlog idea, Fill free slots, Set up our brand or Plan the next 2 weeks. Or just type what you want. You can attach videos, images, audio, logos and fonts.

The agent renders the video with the repository's scripts and checks it. Then deliver_short hands it over: the video folder moves to the project's backlog with its texts, and a draft per account lands on the next free posting slot under Queue › Review.

Image posts work the same way: attach screenshots or ask for a graphic, and the agent puts the images into a folder of their own and hands them over with deliver_short – one image, or a carousel in file-name order (X takes up to 4). Every account that takes images gets a draft; YouTube is left out, and the agent tells you if an image does not fit an account (e.g. Instagram's feed needs 4:5 to 1.91:1).

  1. Type your request, e.g. “Make a new Short about our morning routine”, and press Enter.

  2. Follow the steps in the chat. Cards show each handed-over post and where it is now: waiting for review, scheduled, published or failed.

  3. Open Review from the card or the bar above the input, then approve the draft.

Claude, Codex, Cursor or OpenCode

Every chat runs with one main agent. Settings › Coding agents sets the Main agent for new chats; the agent menu below the input changes it, with model and effort, for a new chat.

Add helper adds the other agents as helpers. They build assets in the background (only Codex makes images) while the main agent keeps going. Write “@codex …”, “@cursor …” or “@opencode …” to talk to a helper directly. Each agent uses your own plan or API key with its provider.

OpenCode runs the models you connected to it – Mistral, Ollama or your own OpenAI-compatible server – so the Studio works with your own LLM too. Pick the model in the agent menu. See Your own LLM.

Runs keep going, Stop ends them

A chat runs on your computer, not in the browser tab. You can switch pages or close the tab: the agent keeps working, and the sidebar lists running sessions under Studio. Notify me sends a desktop notification when it is done.

Stop ends a run; what the agent thought so far is lost. Restarting Publit.Si ends it too. A message sent while the agent works waits and goes out when it is done.

For a Claude Code chat, the header shows the session's command, e.g. claude --resume <id>. Run it in the project repository to continue in your terminal. History lists every chat in the repository – Claude Code, Codex, Cursor and OpenCode from the Studio, plus Claude Code sessions from your terminal – so you can continue them here (just not in both places at once).

Approve commands

Claude Code only: when it needs a command that isn't allowed yet, the chat shows Claude needs your permission with the command and an allow rule. Click Allow (or Allow all), then Continue.

Allowed rules are saved in .claude/settings.local.json of the project, so they also apply in your terminal. The repository's own pnpm scripts run without asking. Codex and Cursor don't ask: they work inside their own sandbox in the project folder. OpenCode doesn't ask either: it gets the same allowed commands as Claude Code, may only write inside the repository, and anything else is refused.

Insights

Studio › Insights holds the project learnings: short points about what works, e.g. “a question as the hook doubles views”. They go into every text the agent and the editor write.

It also shows the Latest weekly analysis. Analyse with AI asks your AI for a new one now. After an analysis, your AI suggests updated learnings; you decide what stays. See Analytics.

Usage

Studio › Usage shows tokens, API value and Shorts per week, Studio and terminal sessions in the project folder, by model, and the most expensive sessions. Enter what your Claude plan costs per month to see what it saved you.

Claude Code only: the numbers are estimated from Claude Code's session files on your computer with Anthropic's API list prices. With a subscription you pay its flat price instead. Codex and Cursor don't report costs; they run on your plan with OpenAI or Cursor. OpenCode reports what its provider charged, if anything.

Check

After deliver_short, the video or images show under Backlog and its drafts under Queue › Review.

Troubleshooting

Windows: "Claude Code (claude.exe) not found"
Install Claude Code with the native installer (step 8); an npm install isn't found.
The agent changes files in my app's repository
Link the project to a separate marketing folder instead: Studio → Change project folder → Create a marketing folder.

Stuck? Search the troubleshooting list or write to us . Leave out secrets and tokens.