Claude Code Skills

DataGenFlow includes built-in Claude Code skills that provide step-by-step guidance for common workflows. Skills activate automatically when Claude Code detects a matching task.

Table of Contents

• Available Skills
• How Skills Work
• Skill Reference
  • Creating Pipeline Templates
  • Testing Pipeline Templates
  • Configuring Models
  • Implementing Blocks
  • Debugging Pipelines
  • Writing E2E Tests

Available Skills

Skill	Triggers When You…
creating-pipeline-templates	Create or modify YAML templates and seed files in lib/templates/
testing-pipeline-templates	Test a pipeline end-to-end, analyze results, or iterate on quality
configuring-models	Set up LLM or embedding models (OpenAI, Anthropic, Gemini, Ollama)
implementing-datagenflow-blocks	Create or modify pipeline blocks in lib/blocks/builtin/
debugging-pipelines	Troubleshoot pipeline failures or unexpected output
writing-e2e-tests	Write Playwright e2e tests for the UI

How Skills Work

Skills live in .claude/skills/ as SKILL.md files. When you describe a task in Claude Code, the matching skill loads automatically and guides the session with:

• Workflows — step-by-step procedures for the task
• Code patterns — correct usage of DataGenFlow APIs and conventions
• Checklists — verification steps before considering the task done
• Common mistakes — known pitfalls and their fixes

No manual activation needed. Just describe what you want to do.

Skill Reference

Creating Pipeline Templates

Location: .claude/skills/creating-pipeline-templates/

Guides through creating YAML pipeline definitions and seed files. Covers:

• Template YAML format and block configuration
• Available blocks (14 total) with inputs/outputs
• Common pipeline patterns (generation, validation, augmentation)
• Seed file format (JSON and Markdown)
• Validation via TemplateRegistry

Example trigger: “Create a pipeline template for generating product reviews”

Testing Pipeline Templates

Location: .claude/skills/testing-pipeline-templates/

Three-phase testing process for pipelines:

1. Dry run — single-seed execution, trace analysis
2. Small batch — 3-5 seeds, consistency and diversity check
3. Quality iteration — identify weak blocks, adjust prompts/config, re-test

Example trigger: “Test the json_generation template end to end”

Configuring Models

Location: .claude/skills/configuring-models/

Provider-specific setup for LLM and embedding models:

• OpenAI — no endpoint needed, sk-... API key
• Anthropic — no endpoint needed, sk-ant-... API key, no embeddings
• Gemini — no endpoint needed, API key from AI Studio
• Ollama — endpoint required (http://localhost:11434), no API key

Includes connection testing, model resolution order, and troubleshooting.

Example trigger: “Set up Ollama as my LLM provider”

Implementing Blocks

Location: .claude/skills/implementing-datagenflow-blocks/

Complete reference for creating pipeline blocks:

• Block structure (BaseBlock, BaseMultiplierBlock)
• UI integration patterns (model dropdowns, enum selectors, template editors)
• LLM and embedding integration via llm_config_manager
• State management and trace_id-keyed caching
• Unit testing patterns

Example trigger: “Create a new block that summarizes text”

Debugging Pipelines

Location: .claude/skills/debugging-pipelines/

Four-phase systematic debugging:

1. Observe — gather evidence, read full errors
2. Trace — follow data flow through blocks
3. Analyze — form testable hypothesis
4. Fix — minimal targeted change, verify

Example trigger: “My pipeline output has unexpected fields”

Writing E2E Tests

Location: .claude/skills/writing-e2e-tests/

Playwright test patterns for the DataGenFlow UI:

• Test file template with sync API and cleanup fixtures
• Navigation, file upload, and modal interaction patterns
• Common UI selectors for all pages
• Integration with run_all_tests.sh

Example trigger: “Write e2e tests for the settings page”