arch-playground/workflow-creator

Workflow Creator

A guided skill that takes you from "I want to automate X" to a working multi-job GitHub Actions workflow using ai-workflow-runner — with AI prompt files, dependency graphs, optional validation scripts, and correct auth setup.

---

When to Use This Skill

Use this skill when you want to:

• Create a new .github/workflows/*.yml that uses ai-workflow-runner for one or more AI steps
• Design a multi-step workflow with parallel and sequential job dependencies
• Generate individual workflows/<workflow-slug>/instructions/<step-name>.md AI prompt files for each step
• Add validation scripts that enforce AI output quality with retry loops
• Edit or extend an existing ai-workflow-runner-based workflow (add/remove/modify steps)
• Get guided help with auth setup (Anthropic API key, GitHub Copilot, custom model)
• Learn how to pass data between AI workflow steps (outputs vs artifacts)
• Evaluate workflow step prompt quality against a sample codebase
• Benchmark step consistency across multiple runs with before/after comparison
• Grade workflow outputs against generated or custom assertions

---

What This Skill Produces

For each workflow you create, the skill generates:

1. .github/workflows/<workflow-name>.yml — the complete multi-job GitHub Actions workflow with correct needs: dependency graph, auth setup, and artifact transfer steps
2. workflows/<workflow-slug>/instructions/<step-name>.md (one per step) — focused AI prompt files with objective, constraints, output format, and success criteria
3. workflows/<workflow-slug>/validation/<step-name>.py (optional, per step) — validation scripts that check AI_LAST_MESSAGE and return retry feedback if the AI output doesn't meet quality criteria
4. .evaluations/workflows/<name>/benchmark.md (when evaluating) — evaluation report with pass rates, failure analysis, delta comparison, and actionable recommendations

---

Prerequisites

• CWD must be your target repository root — the skill writes all files relative to your current working directory
• Node.js / npx (optional but recommended) — used to run action-validator for YAML schema validation after generation. If not installed, validation is skipped with a warning and you'll need to run it manually
• GitHub repository with Actions enabled — to actually run the generated workflow

---

How to Start

To create a new workflow, just say:

• "Create a GitHub Actions workflow"
• "I want to build an AI workflow"
• "Help me set up ai-workflow-runner"

To edit an existing workflow:

• "Edit my workflow at .github/workflows/my-workflow.yml"
• Provide the path to the existing workflow file

To evaluate an existing workflow:

• "Evaluate my workflow"
• "Run eval on service-analysis"
• "Benchmark my workflow steps"

---

Workflow Steps

The skill guides you through 4 focused steps:

| Step | Name | What Happens | | -------- | ---------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------ | | 01 | Discover | Collect workflow name, trigger, action reference, and steps with objectives | | 02 | Dependencies | Map parallel vs sequential relationships, confirm Mermaid diagram, choose auth + data-passing strategy | | 03 | Prompts | Review and edit/accept AI prompt for each step; optionally add validation scripts | | 04 | Generate | Produce all output files, run pre-generation checklist, validate with action-validator | | Evaluate | Run steps against a sample codebase, grade outputs, compare before/after, get benchmark report |

Progress is tracked in .workflow-creator-wip.md at your repo root. If you interrupt the workflow, you can resume from where you left off.

---

Starting the Skill

Read fully and follow: skills/workflow-creator/workflow.md to begin the guided process.