Prd

ML PRD Generator

Create ML-centric PRDs that are evidence-driven, stack-agnostic, and suitable for ML-Ralph.

The Job

1. Receive a project description
2. Ask 3-5 essential clarifying questions (one at a time)
3. Generate a structured ML PRD
4. Save to tasks/prd-[feature-name].md

Important: Do NOT start implementing. Just create the PRD.

Step 1: Clarifying Questions (One at a Time)

Focus on:

• Objective/Metric: What is success? What metric matters?
• Data Context: What data exists? Any leakage risks? Any constraints?
• Evaluation: What validation scheme is appropriate?
• Scope: What should NOT be done?

Example format:

1) What is the primary objective?
   A. Classification
   B. Regression
   C. Ranking
   D. Other: [specify]

Step 2: PRD Structure

Generate the PRD with these sections:

1. Introduction/Overview

Brief description of the ML task and why it matters.

2. Goals

Specific, measurable objectives (bullet list).

3. Assumptions

Explicit assumptions (data availability, metric definitions, constraints).

4. Evaluation Plan

• Metric definition
• Split strategy (random/stratified/group/time)
• Leakage rules

5. User Stories (ML-centric)

Each story must include:

• Title
• Description
• Type: discovery | experiment | evaluation | implementation | ops
• Hypothesis (optional but preferred)
• Evidence Required (what must be logged, including W&B run URL/ID for experiment/evaluation stories)
• Acceptance Criteria (verifiable)

Format:

### US-001: [Title]
**Description:** As a [role], I want [outcome] so that [benefit].
**Type:** discovery | experiment | evaluation | implementation | ops
**Hypothesis:** If ..., then ... because ...
**Evidence Required:** [What must appear in progress.jsonl or artifacts; include W&B run URL/ID for experiment/evaluation stories]

**Acceptance Criteria:**
- [ ] Specific, verifiable criterion
- [ ] Another criterion
- [ ] Ruff check passes
- [ ] Ruff format passes
- [ ] Mypy passes
- [ ] Pytest passes (if tests exist)
- [ ] Evidence logged in progress.jsonl

Important:

• Stories must be small enough for one iteration.
• Acceptance criteria must be verifiable.
• Include evidence logging for every story.

6. Functional Requirements

Numbered list of required behaviors or components.

7. Non-Goals (Out of Scope)

Explicitly list what will not be done.

8. Risks / Uncertainties

Known unknowns and how they’ll be resolved.

9. Success Metrics

Define “done” in measurable terms.

10. Open Questions

Remaining questions that might alter the plan.

ML-Ralph Dynamic Backlog Guidance

PRDs are living documents. ML-Ralph may refine prd.json each iteration based on evidence:

• add/split/reorder/supersede stories
• never delete stories
• log changes in progress.jsonl

Output

• Format: Markdown (.md)
• Location: tasks/
• Filename: prd-[feature-name].md (kebab-case)