markback

---
name: markback
description: "Start Markback in a project, or from one sentence: the person is drawing on their app's screens within a minute, with nothing to paste. Use when someone types /markback, wants to see or review their app, or says what they want built."
argument-hint: "[what you want, if there is no app yet]"
---

The goal is the person drawing on their own app's screens within a minute. Never ask them for an address. Open the review for them, and give them its address only when you cannot open it yourself.

Start from what the person said.

If `review` is not found, use `node <Markback folder>/review.mjs` in its place. The Markback folder is two folders up from this SKILL.md: the one that holds `skills`.

## Set up the New project page

When the person asks to open New project, start Markback with `review start`, then connect their agent to this workspace using the wake instructions in `ideas-and-queue`. Use the workspace name as the review target when there are no screens yet. Run `review open --new --no-open` and open the printed address in the agent’s browser pane. If the agent has no browser pane, use `review open --new` to open it in the person’s browser. The box starts with “Make a todo app.” The person can edit it or press **Send to agent**. Wait for their request before building.

## In their project

Run this in the project's folder:

```bash
review project
```

It finds how the app runs, takes pictures of its main screens, and names the review after the folder. If your desktop session can open the review in its built-in browser, use `review project --no-open` and open the printed URL there yourself. If that fails, run `review open --app "<review name>"` to open their regular browser.

- **"Your app runs with npm run dev, and it is not running":** starting it runs their code. Ask them first, in one line. On a yes, run `review project --yes`.
- **"needs a small Chrome from Google":** that is a download. Ask them first, in one line, saying what it is and how big. On a yes, run the same command again with `--fetch-browser`.
- **"No Chrome or Edge here":** your sandbox has none. Follow "In a sandbox" in the `push-screens` skill: take the screenshots yourself and push them.
- **"There is no app in this folder yet":** make the first screen from what they said (below).

## No app yet

1. Infer what the application needs: its main screen, its real content, what a person does first. Do not ask questions.
2. Write that one screen as a single HTML file, `index.html`, with believable content. No build step, and no back end yet.
3. Run `review project` again. It finds the page and opens the first review.

## Then

Tell them they can draw and press **Send**. From here you keep the review going yourself, and they never run a command: use the `read-marks` skill to wait for their answer and act on it, and the `push-screens` skill after each visible change you make, for as long as this session works on the project.

For a single-screen project, `review project --no-open` can also capture and push the next round. It uses the same review name and screen. Wait for that round in the same session.

Once the first review is open, set up wake as described in the `ideas-and-queue` skill. In Codex Desktop, use the chat's native heartbeat and let the turn end after the first push. Do not register a subprocess or start a blocking wait when the Desktop heartbeat is active.

For Kimi with background wake configured, finish the turn after a push so Markback can resume the selected Kimi Code session on the next request. For other sessions without a Desktop heartbeat, start `review work wait --by "your name"` after each round and do what it prints.

If a command fails or nothing shows, run `review check` and tell the person what it says.