Skip to content
Back to skills

Ralph

ASecurity

Convert an ML PRD into prd.json for ML-Ralph. Use when you have an ML PRD and need prd.json. Triggers on: convert this prd, turn this into ml-ralph format, create prd.json from this, ralph json.

  • 39 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added February 7, 2026
ai-agents

Security analysis

A100/100

Scanned February 12, 2026

npx -y skills add pentoai/ml-ralph --skill ralph --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Ralph?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Ralph
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/pentoai-ralph/badge)](https://www.skillsdirectory.com/skills/pentoai-ralph)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: ralph
description: "Convert an ML PRD into prd.json for ML-Ralph. Use when you have an ML PRD and need prd.json. Triggers on: convert this prd, turn this into ml-ralph format, create prd.json from this, ralph json."
---

# ML-Ralph PRD Converter

Converts ML PRDs into the `prd.json` format used by ML-Ralph.

---

## The Job

Take a PRD (markdown or text) and convert it to `prd.json` in the ML-Ralph directory.

---

## Output Format

```json
{
  "project": "[Project Name]",
  "branchName": "ml-ralph/[feature-name-kebab-case]",
  "description": "[Short description]",
  "userStories": [
    {
      "id": "US-001",
      "title": "[Story title]",
      "description": "As a [role], I want [outcome] so that [benefit].",
      "type": "discovery | experiment | evaluation | implementation | ops",
      "hypothesis": "[Optional hypothesis]",
      "evidenceRequired": "[Required evidence to log]",
      "acceptanceCriteria": [
        "Criterion 1",
        "Criterion 2",
        "Ruff check passes",
        "Ruff format passes",
        "Mypy passes",
        "Pytest passes (if tests exist)",
        "Evidence logged in progress.jsonl"
      ],
      "priority": 1,
      "passes": false,
      "notes": "",
      "supersededBy": "",
      "risk": ""
    }
  ]
}
```

---

## Story Size: The Number One Rule

Each story must be completable in **one iteration**. If you cannot describe the change in 2-3 sentences, split it.

---

## Story Ordering: Dependencies First

Order stories so earlier ones unlock later ones:
1. Discovery/evaluation scaffolding
2. Baseline experiments
3. Improvements and analysis
4. Operationalization

---

## Acceptance Criteria Rules

- Criteria must be verifiable.
- **Always include** Ruff/Mypy/Pytest checks and evidence logging.
- For experiment stories, include the specific metric and logging requirement, including W&B run URL/ID.

---

## Conversion Rules

1. Each user story becomes one JSON entry
2. IDs sequential: US-001, US-002, ...
3. `priority` orders execution
4. `passes: false` for all
5. `supersededBy` empty string for all
6. `branchName` must start with `ml-ralph/`

---

## Dynamic Backlog Note

ML-Ralph refines `prd.json` every iteration based on evidence. This is expected and part of the loop. Do not attempt to “lock” the backlog.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.