Building Connectors with datahub-skills

datahub-skills is a Claude Code plugin that accelerates connector development with specialized skills for planning, standards loading, and code review. It is designed to be used alongside the adding a source guide — the skills handle research, scaffolding, and review; you handle the code.

:::note

This guide assumes you have already set up your local development environment per the developing guide.

:::

Installing datahub-skills

Option 1 — install script (recommended):

curl -fsSL https://raw.githubusercontent.com/acryldata/datahub-skills/main/install.sh | bash

Option 2 — npx:

npx datahub-skills install

After installation, confirm the plugin is registered:

cat ~/.claude/plugins/installed_plugins.json

{
  "version": 2,
  "plugins": {
    "datahub-skills@datahub-skills": [
      {
        "scope": "user",
        "installPath": "/Users/you/.claude/plugins/cache/datahub-skills/datahub-skills/1.2.0",
        "version": "1.2.0"
      }
    ]
  }
}

Available Skills

Workflow

1. Load standards

Run this at the start of every Claude Code session. It loads the golden connector standards into context so all subsequent skills and code generation follow established patterns.

/datahub-skills:load-standards

Expected output:

Loaded connector standards. Ready to review.

Standards loaded:
- main.md — base classes, SDK V2, config patterns
- patterns.md — file organization, error handling, warning/failure reporting
- testing.md — test requirements, golden files, anti-patterns
- code_style.md — type safety, naming conventions, mypy compliance
- containers.md — container hierarchy design
- lineage.md — SqlParsingAggregator, view lineage, query log lineage
...

:::note

Standards are loaded into your current Claude Code context window. If you start a new session, run /load-standards again before continuing.

:::

2. Plan the connector

Run the planning skill before writing any implementation code. It researches the source system, asks you a set of scoping questions, then generates a _PLANNING.md document that serves as the implementation blueprint.

/datahub-skills:datahub-connector-planning build a connector for [source name]

The skill runs in four stages:

1. Classify — identifies the source category (SQL database, BI tool, orchestration, etc.) and asks you to confirm
2. Research — gathers API docs, Python client libraries, similar existing connectors, Docker images for testing
3. Scope — asks you three questions about test environment, permissions, and which features to prioritize
4. Document — generates _PLANNING.md at the repo root with entity mapping, architecture decisions, and an ordered implementation plan

After the document is generated, you will see a summary and a prompt:

## Planning Document Created

Location: `_PLANNING.md`

### Key Decisions:
- Base class: StatefulIngestionSourceBase — no SQLAlchemy dialect exists
- Entity mapping: Pipeline → DataFlow, Table → DataJob, outlet lineage to destination Datasets
- Lineage approach: outlet URNs from state files; inlet URNs user-configured
- Test strategy: filesystem fixtures (no live service required)

Do you approve proceeding to implementation?

Reply approved to proceed, or describe changes and the skill will revise the document before continuing.

3. Build

This is your implementation step. Use the plan in _PLANNING.md as the guide and refer to CLAUDE.md for build, lint, and test commands.

Key commands during development:

# Lint and format Python code
./gradlew :metadata-ingestion:lintFix

# Run unit tests
./gradlew :metadata-ingestion:testQuick

# Run a single test file
pytest tests/unit/myconnector/test_myconnector_source.py -v

See the developing guide for the complete command reference.

4. Review the connector

Run the review skill before opening a PR. It performs a full standards compliance review — architecture, code quality, type safety, error handling, test coverage, and documentation — and produces a verdict with prioritized findings.

/datahub-skills:datahub-connector-pr-review

The skill outputs a structured report:

## Connector Review Report

Verdict: NEEDS CHANGES — 3 blockers require fixes before merge.

| Category          | Status         |
| ----------------- | -------------- |
| Architecture      | ✅ PASS        |
| Code Organization | ✅ PASS        |
| Error Handling    | 🔴 NEEDS FIXES |
| Type Safety       | 🟡 NEEDS FIXES |
| Test Quality      | 🔴 NEEDS FIXES |
| Performance       | ✅ PASS        |

### 🔴 BLOCKERS (Must Fix)

B1 — Silent exception swallow in state.json parsing
  dlt_client.py:269 — except Exception: pass with no logging...

Severity levels:

After fixing blockers, re-run the review skill. It resumes with full context from the previous session:

/datahub-skills:datahub-connector-pr-review

The skill_docs/ convention

Commit the skill-generated artifacts alongside your connector code. This gives PR reviewers full context on design decisions and known issues, and creates a paper trail for follow-up work.

Place them in src/datahub/ingestion/source/<connector>/skill_docs/:

src/datahub/ingestion/source/myconnector/
├── __init__.py
├── config.py
├── myconnector.py
├── myconnector_client.py
├── myconnector_report.py
└── skill_docs/
    ├── _PLANNING.md                              # /datahub-connector-planning output
    └── _datahub-connector-pr-review-YYYY-MM-DD.md  # /datahub-connector-pr-review output

Tips

• Always run /load-standards at the start of a new Claude Code session — standards are not persisted across sessions
• The planning skill will ask you scoping questions before generating _PLANNING.md; answer them rather than skipping, as they determine the architecture decisions in the plan
• The review skill catches issues the checklist cannot — error handling gaps, trivial tests, type safety violations — run it even if your code looks clean
• Commit skill_docs/ alongside the connector code; PR reviewers use it to understand design decisions without re-reading the full diff
• If a review session is interrupted, re-invoke the skill and it will resume from where it left off