Prd to tasks

prd-to-tasks — PRD.md → executable task list

This skill turns a PRD into a flat, ordered list of tasks small enough to be picked up by a developer or an AI executor. The output is .workflow/tasks.md — Markdown with - [ ] checkboxes — designed to be:

• Readable by humans as a project punch-list.
• Importable into beads (the gstack convention), GitHub Issues (one task per issue), or fed to ralph-tui as the task source.
• Pushable to Linear via the linear-scrum skill (Setup mode), which reads tasks.md through the Linear MCP and creates one issue per task — not a CSV import.
• Atomic enough that each line items represents 2–8 hours of work for a typical developer (or one focused session for an AI agent).

When this skill applies

• A .workflow/PRD.md exists and phase is prd_drafted.
• The user asks to "make the issues", "split the PRD", "create the task list", "ticket this".
• The orchestrator routes here from prd_drafted (optional next step before scaffolding).

If .workflow/tasks.md already exists, enter revise mode: read the existing tasks and propose specific edits (add, remove, re-order, re-size) rather than overwriting.

Contract

This skill follows the dev-flow contract — see references/contracts.md for the canonical schema. Key facts:

• Reads <root>/.workflow/PRD.md (mandatory) and .workflow/PROJECT.md (for stack/audience context).
• Writes <root>/.workflow/tasks.md.
• Sets phase = "tasks_split" on success (only if current phase is earlier in the enum).

Workflow

Step 1 — Read inputs and decide structure

Read .workflow/PRD.md. Extract:

• The list of user stories (As a … I want … so that …).
• The acceptance criteria for each story.
• Any non-goals (these become negative tasks: explicit "do NOT do X" notes, not work items).
• Technical constraints (e.g., "must use Postgres" → tasks like "set up Drizzle + Neon" go on the list early).

Group tasks into epics when there are 3+ user stories. With fewer stories, keep a flat list.

Step 2 — Decompose each user story into tasks

For each user story, ask: "What must exist for the acceptance criteria to be true?" Walk down from there:

1. Data layer — DB schemas, migrations, seed data.
2. API / server logic — endpoints, server actions, validation, auth checks.
3. UI — pages/routes, components, states (loading/empty/error).
4. Auth & permissions — who can access this, what guards exist.
5. Telemetry & error handling — logging, error boundaries, user-visible errors.
6. Tests — unit + integration where applicable; explicit task even if "smoke test only".

Skip layers the story doesn't actually need. A purely UI story doesn't need a DB task.

Sizing target: each task ≈ 2–8 hours of focused work. If a task feels bigger, split. If a task is < 1 hour, merge with a sibling.

Title convention: imperative verb first ("Add user-settings page", "Wire login flow", "Set up Drizzle schema for posts"). Avoid "Implement X" — too vague.

Step 3 — Write tasks.md

Use this exact format:

# <Project Name> — Tasks

> Generated by `prd-to-tasks` from PRD.md on <ISO date>. Compatible with beads, GitHub Issues import, `linear-scrum` (via Linear MCP), ralph-tui task source.

## Setup

- [ ] **Set up Next.js + shadcn scaffold** *(addressed by `design-md-to-app`)*
- [ ] **Configure Drizzle + Neon** *(addressed by `module-add db`)*
- [ ] **Wire better-auth + email/password** *(addressed by `module-add auth`)*

## Epic: <user-story name or epic title>

### US-1: As a <persona>, I want <action>, so that <outcome>

- [ ] **<Imperative title>** — <one-sentence body>
  - Acceptance: <bulleted criteria pulled from PRD>
  - Files likely touched: `<best-guess paths>`
  - Estimated: <2-8h>
- [ ] **<Imperative title>** — …

### US-2: ...

…

## Non-goals (do NOT do)

- <Item from PRD non-goals>
- <Item from PRD non-goals>

## Open questions

- <Question pulled from PRD's "Open questions" section>

Rules:

• One ## Epic: heading per epic, one ### US-N: per user story.
• One - [ ] per concrete task. Sub-bullets under it are details, not separate tasks.
• The *(addressed by …)* annotation marks tasks that are owned by another dev-flow skill — the user (or orchestrator) doesn't have to write code for these, just invoke the right skill. Always tag scaffolding/setup tasks with which skill owns them.
• Order: Setup first (foundation), then epics in priority order from the PRD.
• Don't enumerate [ ] checkboxes for sub-bullets — they're details, not actionable items.
• The - [ ] **<Title>** — <body> format (with the em-dash — separator) is load-bearing. linear-scrum/scripts/task_key.py parses each task line with split(" — ")[0] to derive the title it hashes for dedup (meta.json#linear.issue_map). Do not swap the em-dash for a hyphen, colon, or any other separator — doing so breaks re-run idempotency in linear-scrum.

Step 4 — Update state and report

Update .workflow/meta.json:

• if current phase is earlier than tasks_split, set phase = "tasks_split"
• bump updated_at
• append history:

  {
    "skill": "prd-to-tasks",
    "ran_at": "<now>",
    "outputs": ["tasks.md"],
    "phase_before": "<prev>",
    "phase_after": "tasks_split"
  }
  

Tell the user:

• Number of epics + tasks generated.
• Any open questions still flagged.
• Next-step proposal: usually figma-to-design-md (if not yet run and Figma exists) or design-md-to-app (if DESIGN.md is in place), or linear-scrum (Setup) if they want to run the project as agile scrum in Linear.

If the user wants to move tasks into beads/GitHub Issues/Linear, point at the file — it's already in the right shape. They can use gh issue create per task, ralph-tui-create-beads / ralph-tui-create-json skills if installed, or invoke linear-scrum (Setup mode) to push tasks.md into Linear via MCP — not a CSV import.

Important constraints

• Don't invent acceptance criteria. If the PRD doesn't specify them for a story, ask the user before writing tasks for that story. A task with vague acceptance is worse than no task.
• Don't generate a 100-task list from a 3-story MVP. If the math gives you more than ~15 tasks, you're over-decomposing. Merge.
• Don't include tasks for non-goals. They're explicitly out of scope. List them under ## Non-goals so they stay visible but not actionable.
• Tag setup tasks with the owning skill. "Set up auth" should always say *(addressed by module-add auth)* — this is the bridge between the task list and the skill ecosystem.
• One file, one project. Don't shard tasks.md across multiple files.