01 · Quickstart
Five minutes to your first task.
Three steps, no extra setup. Each one shows exactly what you see in the product.
Create a plan from a brief
Log in, open the New Project page, write a short brief. Example: a cashier app for a coffee shop. The AI assembles the structure, PRD, and tasks until the plan status is ready.
New mission
What should we build?
a cashier app for a coffee shop, with tables, orders, and daily reports
Copy the prompt, paste it into the agent
Click Start implementation on the plan page, copy the prompt that appears, paste it into your AI agent. The agent connects to the plan, reads the PRD, then works on its own.
Mission prompt
Start implementation$ scratch-agent plan get rv_8f3a
✓ plan ready: 3 phases, 28 tasks
› token rv_•••••••• already included
Paste into OpenCode, Claude Code, or Cursor. The agent reads the PRD and then works through the tasks one by one.
Watch the live progress
The task currently in progress shows a spinner, completed ones get struck through, and the progress bar moves in real time. If a task fails or hits a checkpoint, the agent stops and waits for your call.
Live progress
polls every 5 seconds02 · Concepts
How Scratch Agent thinks.
Every plan follows the same structure. Once you know the pipeline, reading any plan takes seconds.
Plan
brief + PRD
Feature
phase = feature
Sub-feature
problem unit
Task
F01-S02-T03
Every task has a unique ref like F01-S02-T03: phase 1, sub-feature 2, task 3. The execution order is decided by the server, not the agent.
Plan structure
Plan → Feature → Sub-feature → Task. Every task has a unique ref like F01-S02-T03.
Phase
One feature = one phase. The agent finishes every task in that phase first before moving on, and the web progress moves with it.
Layer
Tasks have a frontend, backend, or qa layer. Within one phase, the natural order is frontend → backend → qa.
Dependencies
A task can depend on another task. The server always picks a task whose dependencies are done, so the agent is never out of order.
Checkpoint
When switching phase or layer, the agent is asked to stop and report first. You verify, then tell it to continue.
Status
Task: pending → in_progress → done | failed. Plan: generating → ready → implementing → done. A failed task blocks the plan until retried.
03 · Prompt Agent
One prompt, the agent gets to work.
Every plan has a mission prompt generated for you, with the plan identity and full instructions already inside. No commands to memorize, nothing to write yourself.
# mission sent once, the agent loops on its own
$ scratch-agent task next --plan futsalgo
{ "ref": "F01-S01-T02", "blocked": false, "checkpoint": false }
$ scratch-agent task start F01-S01-T02
...the agent works on this task...
$ scratch-agent task complete F01-S01-T02
{ "done": false, "next": "F01-S01-T03" }
task failed → the plan is blocked, the agent stops and reports.
checkpoint → the agent waits for your verification before continuing.
How to get it
Open a plan that is ready, click the Start implementation button, and copy the prompt that appears. Your access token is already included in it.
What the prompt contains
Instructions to connect to the plan, read the PRD, then loop: pick up the next task → do the work → mark it done. Order and dependencies are managed by the server.
Supported agents
OpenCode, Claude Code, Cursor, Codex, and any other agent that accepts text instructions. Paste the prompt and let the agent work.
Safety mechanism
If a task fails, the plan is blocked and the agent stops and reports. At a checkpoint, the agent waits for your verification before continuing.
Ready to turn a brief into a plan?
Create your first plan now with the Free plan: 3 generates per 24 hours, no credit card.
Start with the Free plan