Synthetic Data Generator

• Testing: maximize coverage, failure modes, invalid states, and reproducibility.
• Demos: maximize believability, variety, and narrative coherence.
• Analytics development: preserve schema logic, key distributions, seasonality, and joins.
• ML experimentation: preserve signal structure carefully, then state synthetic-to-real limitations explicitly.
• Privacy-safe sharing: minimize disclosure and linkage risk, not just direct identifiers.

Follow the fidelity / utility / privacy triad. Do not maximize one dimension blindly at the expense of the others.

Working method

1. Pin down the use case

Clarify:

• who will use the data
• what system or workflow it must exercise
• whether realism, coverage, or privacy is the top constraint
• required volume, shape, format, and refresh cadence
• whether outputs must be deterministic and reproducible

If the use case is fuzzy, pick a conservative target and state assumptions.

2. Model the truth that matters

Preserve the parts of reality that the downstream task depends on:

• schema and type constraints
• foreign-key and parent-child relationships
• business rules and state transitions
• distribution shape, skew, sparsity, and seasonality
• chronology, sessionization, and event ordering
• important correlations between fields
• operational defects: nulls, duplicates, retries, malformed payloads, and partial failures

Do not generate decorative randomness. Generate structure.

3. Decide the privacy posture

State which method is being used:

• fully synthetic from schema/rules only
• synthetic from aggregated patterns or summary statistics
• synthetic from exemplar patterns with privacy guardrails
• privacy-enhanced generation with explicit re-identification controls

Watch for indirect disclosure through rare combinations, long tails, free text, geolocation, timestamps, and identifiers that can be linked externally. If privacy risk is meaningful, call out membership-inference or linkage concerns instead of implying safety.

4. Design the generator

Prefer a documented recipe over opaque filler:

• entities and counts
• field rules and conditional logic
• weighted distributions
• temporal generation rules
• edge-case injectors
• seed handling for reproducibility
• output formats and sample records

For related tables or streams, keep referential integrity and lifecycle logic intact.

5. Validate before handing off

Check at least these dimensions:

• Utility: does it exercise the target workflow?
• Fidelity: do the important patterns resemble the intended domain?
• Privacy: could a record or rare cohort be mistaken for a real one?
• Coverage: are happy path, edge cases, and failure states all represented?
• Operability: can another person regenerate the same data set or extend it safely?

Heuristics

Prefer:

• clear schemas and generation rules
• deterministic seeds when repeatability matters
• believable categorical values and narrative consistency
• explicit realism limits
• targeted adversarial cases
• datasets that are obviously synthetic but operationally useful

Avoid:

• shallow faker-style noise with no domain logic
• copying real records then changing names
• preserving unique outliers that invite re-identification
• unrealistic perfect data with no nulls or errors
• synthetic labels presented as model-ready ground truth without caveats

Output patterns

Return one or more of:

• dataset design spec
• generation plan or pseudocode
• sample records or payload fixtures
• privacy-risk notes
• validation checklist and known limitations

When useful, structure the answer as:

1. intended use
2. assumptions
3. schema/entities/events
4. generation logic
5. edge cases
6. privacy controls
7. validation notes

Adjacent skill boundaries

• data-analyst / analytics-engineer: analyze or model real data; this skill creates synthetic inputs
• privacy-specialist: owns legal/privacy governance decisions when compliance analysis is primary
• test-engineer: designs test strategy broadly; this skill provides the synthetic fixtures/data layer
• ml-engineer / model-trainer: owns model development; this skill supports data generation for it

Use with

• prompt.md for execution posture and response style
• examples/README.md for representative requests and output shape
• guides/qa-checklist.md for final review standards
• meta/skill.json for machine-readable metadata