| name | synth-data |
| description | Generate synthetic tabular datasets (CSV/JSON) for demos, tests, and dashboards — from a bundled business preset, a natural-language description, or a reference document. Validate generated data against its schema. Use when the user asks for fake/test/sample/demo/dummy data, a synthetic dataset, "rows of customer/order/patient data", to mock a database table, or to turn a business description or a reference PDF into a reusable data schema. Wraps the self-contained `pocsynth` CLI; the generation half is free and offline.
|
synth-data
Generate believable synthetic tabular data (not documents). This skill
drives the structured-data half of the pocsynth CLI. The generation and
validation steps are free and offline (Faker); only schema inference from a
reference document or a description spends Bedrock tokens.
This is the data-generation companion to the pocsynth skill (which converts
PDFs into synthetic documents). If the user wants a rewritten/redacted
document, use pocsynth. If they want rows of data, use this skill.
The CLI
Use the same bundled script the pocsynth skill ships:
./skills/pocsynth/pocsynth.py --json <subcommand> ...
It self-deploys via uv run --script (no pip install). Always pass --json
when driving it from the skill; --json comes BEFORE the subcommand.
Decision tree (pick the seed source)
Does the user have a source artifact?
├─ No, and a common domain fits → PRESET (free, instant) Recipe A
├─ No, but a specific business → FROM-PROMPT (one small paid call) Recipe C
└─ Yes, a real sample PDF → FULL PIPELINE (paid once) Recipe D
Always finish by VALIDATING the data (Recipe B).
Cost model
| Step | Spends Bedrock? | Notes |
|---|
presets, generate, test | No | Free, offline, unlimited. No doctor needed. |
schema --from-schema (lint) | No | Offline schema lint / document / --fix. |
schema --from-prompt | Yes (small) | One call; typically pennies. |
extract, schema --from-sample | Yes | Run doctor first; gate with estimate --for. |
Cost gate: before extract, run ./skills/pocsynth/pocsynth.py --json estimate <pdf> --for extract. If result.total_cost_usd > $0.10, surface it
via AskUserQuestion before proceeding. schema --from-prompt is pennies — just
mention it spends a little.
Recipes
Recipe A — Preset dataset (free)
... --json presets to list b2b_saas, ecommerce_orders, healthcare_lite.... --json generate --preset <name> --rows <N> --seed <S> --format csv -o ./out- Validate (Recipe B). Use
--seed whenever reproducibility matters.
Recipe B — Validate (free; always do this)
... --json test --rows ./out/rows.csv --schema ./out/schema.json
Exit 0 ⇒ valid. Exit 7 (DATA_INVALID) ⇒ read
result.context.report.violations (row/field/rule) and surface a summary. Do
not retry on exit 7 — the data or schema needs fixing.
Recipe C — Describe a business → schema → data (one small paid call)
doctor if a fresh env.... --json schema --from-prompt "a B2B SaaS company's customer accounts with plan tier and MRR" -o ./out- Review
./out/schema.md (the data dictionary), then generate (A) + test (B).
Recipe D — Real PDF → reusable schema → unlimited free rows (the cost-saver)
doctor if a fresh env.... --json estimate report.pdf --for extract → cost gate (AskUserQuestion if >$0.10).... --json extract report.pdf -o ./out (paid; PII-audited by default).... --json schema --from-sample ./out/sample.json -o ./out (paid). Review
./out/schema.md and ./out/lint_report.json.... --json generate --schema ./out/schema.json --rows 10000 --seed 1 (free) + test (B).
PII — the guarantee to communicate
extract audits the values it pulls (Amazon Comprehend, on by default), and
schema --infer never lets a real PII value become an enum — PII fields
are bound to Faker providers and the real values are discarded. So:
- ✅ the generated dataset is safe to share;
- ⚠️ the extract sample (
./out/sample.json) and the PII-audit CSV
contain real values — treat them as sensitive, don't hand them out.
When the source PDF holds real customer data, do not pass --no-pii-audit.
Exit codes
Same contract as the pocsynth skill (see its SKILL.md). The one addition:
exit 7 DATA_INVALID from test — surface the violation report, don't retry.
When NOT to use this skill
- The user wants a rewritten/redacted document (HTML/Markdown) → use
pocsynth.
- The user wants to scan an existing file for PII →
pocsynth pii-audit.
