← All guides
Prompt · Sandbox Testing · Reusable

Sandbox Seed
realistic test data, on demand

The hardest part of testing a new feature is never the feature — it's getting enough realistic records into the sandbox to see it actually work. This is the prompt that does it for you. Not a skill, just a very good prompt structure you fill in and run in Claude Code.

Updated Jul 10, 2026
Why a prompt, not a skill Seeding data changes every time — different object, different states, different volume. A fixed skill would fight that. A strong prompt structure flexes to whatever you're testing while still enforcing the rules that keep it safe: fake data only, relationships resolved, a real spread of states.
The prompt

Fill in the [BRACKETED] parts, paste it into Claude Code with your sandbox connected, and let it confirm your schema before it loads anything.

sandbox-seed.prompt
I have a new feature in my Salesforce sandbox and I need realistic test data so I can see it
working across every state — without building records by hand.

THE FEATURE
[ONE OR TWO SENTENCES — what the feature does and what a user needs to see it working]

WHAT TO SEED
- Object(s): [LIST THE OBJECTS THIS FEATURE READS OR WRITES]
- How many records: [e.g. 25 — enough to feel real, not so many it's slow]
- States I need to see: [LIST THE VARIATIONS — e.g. active vs inactive, each status value,
  records with and without a related child, edge cases the UI must handle]

RULES — follow all of these
1. Generate realistic but clearly FAKE values. No real names, no production data, nothing that
   could be mistaken for a real person or record.
2. Respect every required field and field type. If a field is a picklist, use only valid values.
3. Create parent records BEFORE child records so every relationship resolves.
4. Give me a spread across the states listed above — not 25 copies of the happy path.
5. Load it into my sandbox using the Salesforce CLI, then tell me exactly what you created
   (counts per object) so I can verify before I test.

BEFORE YOU START
Confirm my schema with me first — which objects, which required fields, which relationships.
Do not assume my field names. Ask me anything you need, then show me the plan before you load it.
See it in action

A real run of this prompt — confirming the schema, then loading a spread of fake records into a dev sandbox.

The three rules that matter
RuleWhy
Fake data onlyNever seed a sandbox from real or production records. In a regulated org especially, that's a privacy problem, not a shortcut.
Parents before childrenIf a child record points at a parent that doesn't exist yet, the load fails. Order matters.
A spread of statesTwenty-five happy-path records prove nothing. You need the empty, the edge, and the unusual to actually test the UI.
Confirm before you load The prompt ends by telling Claude to confirm your schema and show the plan first. Keep that line. It's the checkpoint that stops a wrong-shaped load — and it doubles as the moment to make sure you're comfortable sharing your object structure with an AI tool.
Found this useful?