Claude Data Annotation Plugin

April 27, 2026 · View on GitHub

End-to-end toolkit for turning raw data into an annotated dataset and publishing it to Hugging Face. Covers the full lifecycle: ingest, profile, clean (PII, columns, format), design an annotation schema, annotate (interactively with Claude or via Gemini batch inference), review, and publish.

The plugin is built around a small set of orchestrator skills delegating to specialized subagents — instead of one micro-skill per micro-operation, the orchestrator looks at the data, talks to the user, and figures out which steps are needed.

Skills

  • shape-dataset — top-level prep workflow. Ingests a source, profiles it, proposes a prep plan (PII, columns, format, splits, schema), executes approved steps. Hands off to annotation or publication.
  • annotate-with-claude — Claude annotates a small dataset interactively in-session against a locked schema. For runs where Label Studio or batch inference would be overkill (typically tens to low hundreds of records).
  • scaffold-annotation-env — generates a workspace for large-scale annotation via the Gemini batch inference API, with Python boilerplate (run, poll, validate).
  • hf-setup — creates a Hugging Face dataset repo (asks public/private), copies the prepared data over, generates the dataset card, pushes via huggingface-cli.
  • ingest-source — stages raw data from GitHub, a local path, or a remote URL into a known working directory.
  • init-dataset-repo — initializes a local git repo with the conventional HF dataset layout, LFS rules, license, and card stub.

Subagents

The orchestrator skills delegate to these — they are not invoked directly by the user.

  • data-profiler — schema inference, stats, encoding, samples; flags concerns.
  • pii-scanner — detects PII (direct and quasi), proposes redaction strategy per column; can apply on approval.
  • column-curator — recommends keep/drop/rename/recast/derive per column for the target task.
  • schema-designer — proposes annotation schema and guidelines from data + task; iterates with the user before locking.
  • format-normalizer — executes format conversions, encoding fixes, JSON flattening, reshape-for-annotation, and splits.
  • review-annotations — audits finished annotations: schema validation, label distribution, sampled spot-checks, agreement analysis.

Typical workflow

ingest-source
   ↓
shape-dataset  ──→  data-profiler, pii-scanner, column-curator, format-normalizer, schema-designer
   ↓
annotate-with-claude   (small)        OR        scaffold-annotation-env   (large)
   ↓                                                ↓
review-annotations                               review-annotations
   ↓                                                ↓
hf-setup                                         hf-setup

Installation

claude plugins marketplace update danielrosehill
claude plugins install data-annotation@danielrosehill

After installing, restart Claude Code.

Requirements

  • huggingface-cli (logged in) for hf-setup.
  • GEMINI_API_KEY in the annotation workspace .env for scaffold-annotation-env.
  • Python with pandas, pyarrow, presidio-analyzer, jsonschema (skill scripts install via uv on first run).

Data storage

  • Plugin-managed working state: ${CLAUDE_USER_DATA:-${XDG_DATA_HOME:-$HOME/.local/share}/claude-plugins}/data-annotation/.
  • User-owned dataset repos: a path the user picks during onboarding (typically under ~/repos/ or ~/Documents/); only a pointer is stored under $CLAUDE_USER_DATA.

License

MIT.