Loading page…
The shortest true path from an idea to an agent writing code against it. Five minutes, one plan, one terminal.
Sign in, then start a new plan and describe what you want to build in a paragraph or two. Plain prose is fine — the wizard asks for the rest.
The wizard walks four steps: your idea, the structure it proposes, the PRD it drafts, and the task board it derives from that PRD. You approve each step before the next one runs. Nothing is generated behind your back, and you can revise any step in the chat panel before moving on.
The PRD is the contract everything downstream is generated from — the task board, the estimates, and every specification an agent will read. It is worth ten minutes now.
Look for the things a model cannot know: the constraint that rules out an approach, the integration you already have, the thing that is explicitly out of scope. Say them in the chat panel and let the PRD be revised. A vague PRD produces vague tasks, and an agent will implement the vagueness faithfully.
When the task board is generated the plan reaches tasks_ready. Each task carries its own
acceptance criteria, a test plan, and its position in a phase and a layer — that ordering is what
lets an agent work without asking you what comes next.
This is the point where you decide how much rope the agent gets. By default it stops at every phase and layer boundary and waits for you. In the plan's settings you can turn on autonomous mode and it will run the board end to end instead. The kill switch is never gated: you can turn it off at any time, on any plan.
In the repository you want the work to happen in:
npx instantprd login
npx instantprd init --plan <planId>
login opens a browser once per machine and stores a token. In CI, pass --token instead.
init writes the agent configuration pack for the CLIs you use — the reviewed skill file, an
AGENTS.md marker block, and .instantprd.json with the plan id and the app URL. The plan id is
in your plan's URL, and the workspace shows the whole command with it filled in.
npx instantprd task next --json
npx instantprd task start T-014
npx instantprd task complete T-014 --evidence evidence.json
task next returns the full specification for the next task in order: description, acceptance
criteria, test plan, the files it touches, and whether a checkpoint is waiting. task start claims
it. task complete submits it with the evidence the task asked for.
A task that cannot be finished is reported, not skipped:
npx instantprd task fail T-014 --reason "blocked by a missing API key"
Exit codes are meant to be branched on: 0 ok, 1 error, 2 authentication required,
3 checkpoint required, 4 evidence required or invalid.
The board updates live as the agent works. When it reaches a checkpoint it stops and the board says
so; press Continue and the agent's next task next proceeds. Evidence arrives attached to the task
that submitted it, and you either accept it or send the task back.
Read Agent handoff for the full command surface, the checkpoint protocol and the evidence format. If your agent speaks MCP rather than a shell, read MCP server instead. Public API covers reading your plans programmatically.