ViggyV/pipeline-creator

Pipeline Creator

I'm using the pipeline-creator skill to create a custom agent pipeline.

Process

1. Understand the Workflow

Ask the user:

• What is the end goal? (e.g., "build a SaaS app", "create training materials")
• Which agent should own this? (coding-agent, docs-agent, startup-agent, or a new one)
• What stages are needed? List the steps from start to finish

2. Map Stages to Skills

For each stage, identify the right skill from the library. Read QUICK_REFERENCE.md to find matching skills.

Common skill paths:

obra-superpowers/skills/brainstorming
obra-superpowers/skills/writing-plans
obra-superpowers/skills/requesting-code-review
obra-superpowers/skills/verification-before-completion
obra-superpowers/skills/finishing-a-development-branch
backend-api/api-designer
backend-api/database-schema
frameworks/fastapi-builder
frameworks/react-component
devops-infrastructure/docker-composer
code-quality/code-reviewer
code-quality/test-generator
code-quality/security-auditor
business-communication/pitch-coach
business-communication/email-polisher
anthropics-official/document-skills/pptx
data-engineering/analytics-builder
community/d3js-visualization

3. Determine Stage Types

For each stage, choose:

• skill — single skill invocation
• skill-chain — multiple skills in sequence (same context)
• parallel — independent branches dispatched as subagents

Parallel rule: Branches MUST be independent. No shared files. No shared state. If work depends on each other, use sequential stages or a skill-chain instead.

4. Design Data Flow

For each stage:

• Outputs: What files does this stage produce? Use template vars: {date}, {name}, {project}
• Inputs: What does this stage need from earlier stages? Use interpolation: ${stage_id.outputs.key}

5. Add Gates

Decide where to add checkpoints:

• approval — user reviews before continuing (skipped in autonomous mode)
• quality — code review subagent runs (always enforced)
• verification — evidence-based check (always enforced)

Best practices:

• Add approval gates after design/planning stages
• Add quality gates after build/parallel stages
• Add verification gates before the final stage

6. Generate the JSON

Create the pipeline file at:

.claude/skills/agents/<agent>/references/pipelines/<name>.json

Validate the JSON:

python3 -c "import json; json.load(open('path/to/pipeline.json')); print('valid')"

7. Update the Agent

Add the new pipeline to the agent's SKILL.md routing table:

• Add a row to the "Available Pipelines" table
• Add a routing case in the "Route" section

8. Validate

Check against these rules:

• [ ] All id fields are unique
• [ ] ${...} interpolation only references earlier stages
• [ ] Parallel branches have no shared output paths
• [ ] At least one stage exists
• [ ] All skill paths match existing skills in .claude/skills/
• [ ] JSON is valid (no trailing commas, proper quoting)

Pipeline Template

{
  "name": "my-pipeline",
  "description": "What this pipeline produces",
  "mode": "interactive",
  "stages": [
    {
      "id": "stage-1",
      "type": "skill",
      "skill": "category/skill-name",
      "description": "What this stage does",
      "outputs": {
        "output_key": "docs/{date}-{name}-output.md"
      },
      "gate": {
        "type": "approval",
        "prompt": "Stage 1 complete. Continue?"
      }
    },
    {
      "id": "stage-2",
      "type": "skill",
      "skill": "category/another-skill",
      "description": "Next stage",
      "inputs": {
        "input_key": "${stage-1.outputs.output_key}"
      },
      "outputs": {
        "final_output": "output/{name}-result.md"
      }
    }
  ]
}

Integration

• Invoked by: User who wants to create a custom pipeline
• References: pipeline-orchestrator/references/pipeline-schema.md for full schema
• References: pipeline-orchestrator/references/data-contracts.md for data flow rules