data-designer

Use when the user wants to create a dataset, generate synthetic data, or build a data generation pipeline.

2,559stars287forksUpdated 7/17/2026

#synthetic-data #machine-learning #dataset #pipeline #data-generation #python

Security Assessment

Safe(100/100)

Security Score100/100

About data-designer

data-designer helps users create synthetic datasets and build data generation pipelines using the Data Designer library. It is triggered when a user wants to create a dataset, generate synthetic data, or build a data generation pipeline, and it produces a Python configuration script that describes the dataset to generate. Rather than exploring the workspace up front, the skill relies on a Learn step within its chosen workflow to gather everything needed.

The skill runs in one of two modes. Autopilot mode is used when the user implies they don't want to answer questions (e.g., "you decide", "just build it", "surprise me"); otherwise Interactive mode is the default. Each mode has its own workflow file (workflows/autopilot.md or workflows/interactive.md) that is read and followed. The output is a Python file placed in the current directory with a load_config_builder() function returning a DataDesignerConfigBuilder, using PEP 723 inline script metadata for dependencies. It supports sampler and validation columns (which need both a type and params, e.g., sampler_type="category" with dd.CategorySamplerParams), Jinja2 templates in prompt, system_prompt, and expr fields referencing columns like {{ column_name }}, Pydantic models for structured output, custom column generators via @dd.custom_column_generator, and LLM judge columns (LLMJudgeColumnConfig) whose numeric score is accessed with the .score attribute. References cover seed datasets and person sampling, used only when applicable.

Target users are anyone generating synthetic data for testing, prototyping, or ML. By default all columns are kept in the output, seed datasets are only used when explicitly provided, and person data triggers reading references/person-sampling.md. Requirements and limitations include needing the data-designer CLI installed (Python >= 3.10); if it is not found the skill asks permission before installing anything, and network errors during preview may indicate a sandbox blocking outbound requests, which the skill handles by asking to retry with the sandbox disabled.

FAQ

When does it use Autopilot versus Interactive mode?

Autopilot mode is used when the user implies they don't want to answer questions, e.g., saying "be opinionated", "you decide", "make reasonable assumptions", "just build it", or "surprise me". Otherwise Interactive mode is the default. Only the workflow file matching the selected mode is read and followed.

What does the skill produce?

It writes a descriptively named Python file to the current directory containing a load_config_builder() function that returns a DataDesignerConfigBuilder, using PEP 723 inline metadata for dependencies (always including data-designer).

What are the installation requirements?

The data-designer CLI must be installed and requires Python >= 3.10. If the CLI is not found, the skill informs the user and asks whether to create a virtual environment and install it, and it does not install anything without permission.

How do I get the numeric value from an LLM judge column?

LLMJudgeColumnConfig produces a nested dict where each score name maps to {reasoning: str, score: int}. Use the .score attribute to get the numeric value, e.g., {{ quality.correctness.score }}; using {{ quality.correctness }} returns the full dict instead.

Are seed datasets used automatically?

No. Seed datasets are only used when the user explicitly provides seed data or asks to build from existing records, in which case references/seed-datasets.md is read. The skill does not suggest or ask about seed datasets otherwise.