Synthetic Data Design Playbook

Design a synthetic dataset for evals or fine-tuning without shipping a contaminated, low-diversity corpus: plan the generation strategy and schema, catch train/eval/test leakage before it poisons your numbers, score coverage and diversity against explicit anchors, and run a go/no-go validation gate before the data is used. Works in any chat model — no connected tools needed.

A 4-step agentic workflow pack for coding built to run with ChatGPT, Claude, and Gemini. Open the Markdown files, fill the variables, and paste into your model. Most buyers get a reviewable result in about 45 minutes.

• Turn a vague 'we need synthetic data' ask into a concrete generation plan: record schema, source mix, generation methods, difficulty spread, and a train/val/test split policy
• Catch the leakage that silently inflates eval scores — seed reuse, near-duplicates, benchmark contamination, and temporal leakage — with a scored contamination register and a decontamination protocol
• Score your generated sample on coverage and diversity against explicit 1-5 anchors, catch mode collapse and redundancy, and get a targeted rebalancing directive
• Run a real acceptance gate before the dataset is used: a scored Accept / Accept-with-fixes / Reject verdict with hard blocking gates for contamination and label quality
• Encodes practitioner judgment — leakage vectors, coverage dimensions, and scoring weights ML engineers actually use — not a generic 'make me some data' ask
• Every prompt hands its output to the next, so the schema, contamination clearance, and diversity scorecard all feed one final go/no-go decision
• Works in any chat model (ChatGPT, Claude, Gemini) with no connected tools — paste your task spec, sources, and sample records, get structured design artifacts back