Sign inGet started
InstantPRD · Flight directorEnglish

Give your coding agent a spec, not a vibe.

Describe what you want to build, in your own words. In under two minutes you get a full written spec and a task board — so your coding agent builds the thing you meant, instead of guessing.

0/2000

Free account, no card. The paragraph is kept while you sign in.

Start freeSee a finished plan
Built for
Claude Code · Cursor · Codex
Handoff
MCP · CLI
Output
PRD · task board
Idea to board
under 2 min

One paragraph in. Here is some of what came out.

A real run, not a mock-up. Three of the things InstantPRD wrote from the idea it was given.

Demo plan · Bookly
Idea, as typed

A booking app for small salons. Customers open a public page, pick a service, a staff member and a free slotF-01, confirm with their name and phone§6, and can cancel or reschedule from a linkT-005 later. Owners manage services, staff and hours, and get fewer no-shows.

  1. F-01FeaturePhase 1

    Public booking flow

    • Salon page with service catalog
    • Staff & time-slot picker
    • Customer details & confirmation
    • Manage booking via secure link
  2. §6PRD section

    Database schema

    • salon ─< staff · service · customer
    • customer ─< booking >─ staff, service
    • booking ─< notification

    ER diagram + per-table columns

  3. T-005Task

    Build manage booking page (cancel and reschedule)

    frontend · 3 h · /manage/[token] · after T-002, T-004

    • Page shows booking status, service, staff, date and time for a valid token
    • Cancel opens a confirmation dialog and on confirm shows the cancelled state

    +3 acceptance criteria · 2 test steps

Handoff
$ claude mcp add instantprd -e INSTANTPRD_TOKEN=… -- npx instantprd-mcp
› get_next_task → T-005 · Build manage booking page
Checkpointstart_task · 409

Your agent cannot skip ahead. You decide when it moves on.

It finished everything in phase 1 and asked to start the next thing. InstantPRD stopped it, because a phase boundary is yours to clear. Poll the room and see what happens.

  • frontendStandby
  • backendStandby
  • dbStandby
  • integrationStandby

This is the real rule, not a demo: start_task returns 409 at every phase and layer boundary until a human clears it, and complete_task returns 422 with the exact shape the agent must send when evidence is missing.

Flight planMission elapsed T+1:55

What happens in the two minutes

A typical run. Every step streams in, and you can edit its result before the next one starts.

  1. typical T+0:00clarify

    Five questions about scope, users and stack. Answer with one-click chips or skip.

  2. typical T+0:20structure

    Three to six features with sub-features, placed in up to three release phases on a canvas you can rename, rephase and prune.

  3. typical T+0:50prd

    Nine sections — overview, requirements, features, user flow, architecture and ER diagrams, tech stack, acceptance criteria, out of scope.

  4. typical T+1:30specs

    Every task gets a description, acceptance criteria, a test plan, a layer, a page and an estimate — ordered UI first, then data, API, integration.

  5. typical T+1:55handoff

    One command connects Claude Code, Cursor or Codex. The agent pulls the next task, reports evidence, and the board updates as it works.

What the agent actually receives

A plan is only useful if something can build from it. Everything in InstantPRD exists to make the spec complete before the first line of code.

$ instantprd task next --json
{
  "task": {
    "ref": "T-005",
    "title": "Build manage booking page (cancel and reschedule)",
    "layer": "frontend", "page": "/manage/[token]",
    "estimate_hours": 3, "depends_on": ["T-002", "T-004"],
    "acceptance_criteria": [
      { "text": "Page shows booking status, service, staff, da…", "done": false },
      { "text": "Cancel opens a confirmation dialog and on con…", "done": false },
      …
    ],
    "test_plan": ["Playwright: …", "Playwright: …"],
    "feature": { "ref": "F-01", "phase": 1 }
  },
  "progress": {
    "phase": { "current": 1, "total": 3 },
    "layer": "frontend", "completed": 4, "remaining": 34
  },
  "checkpoint": { "required": false }
}

What `get_next_task` returns to the agent — the same object the board shows you.

  1. Acceptance criteria and a test plan on every task

    acceptance_criteria · test_plan

    Not titles. Each task says what done means and how to prove it, so the agent stops guessing.

  2. Layer, page, estimate and dependencies

    layer · page · estimate_hours · depends_on

    Tasks are ordered the way agents build well: UI on mock data first, then schema, API, integration.

  3. A nine-section PRD with diagrams

    get_plan_prd

    Architecture and ER diagrams in Mermaid, explicit out-of-scope, acceptance criteria per feature.

  4. One command to connect

    get_next_task · complete_task

    MCP server and CLI. get_next_task, start, complete with evidence, fail with a reason — status syncs back to the board in seconds.

  5. Revisions with a diff preview

    prd.version

    Ask for changes in chat with @mentions, see the diff, apply. Every revision is a new PRD version you can compare.

  6. Export for your stack

    CLAUDE.md · AGENTS.md

    Markdown, or a ZIP with CLAUDE.md, AGENTS.md, .cursorrules and the InstantPRD agent skill written for your repo.

Start with the paragraph you already have.

Two plans a month, free. No card. English and Bahasa Indonesia.

Create your first plan