Prompt Engineering

Create Reliable JSON Responses

Getting JSON once is easy; getting the same JSON shape on run 500 is the real problem. The consistency mechanics: stable schema, null discipline, and type pinning.

Overview

The painful JSON failures aren't day-one failures — they're run-500 failures: the model that returned phone as a string all week suddenly returns a number, omits an optional key instead of nulling it, or adds a helpful extra field that breaks a strict deserializer. Reliability across runs comes from pinning every degree of freedom: explicit types per field, null-not-omit discipline, no-extra-fields rules, and an example that demonstrates all three. This resource loads a CRM-entry contract tuned for repeated automated calls — every degree of freedom pinned.

Workflow

Pin the formats in descriptions
call_date says YYYY-MM-DD — date format drift is the most common run-500 failure, and the description is where you pin it.
Keep optional fields nullable, never omitted
The null-not-omit rule keeps the key set identical across runs — strict deserializers depend on it.
Ban extra fields explicitly
'No extra fields' stops the model from helpfully enriching your schema on creative days.

Why This Works

Consistency failures are degree-of-freedom failures — the contract pins each one explicitly
A stable key set (null over omission) makes every response structurally identical
Format pinning inside descriptions catches the drift that type declarations alone miss

Best for

Recurring automated calls where every response must deserialize identically
Strictly-typed consumers (C#, Go, TypeScript with strict parsing)
Pipelines where a single shape change pages someone at night

Not for

One-off JSON requests — day-one reliability needs less machinery
Catching drift after it happens — that's the AI Output Validator

Use cases

Stabilizing a prompt that runs hundreds of times a day in an automation
Eliminating the run-500 type drift that breaks strict deserializers
Defining team-wide JSON conventions one contract at a time

Customize This Resource

Opens this setup in JSON Output Prompt Builder. Generate to get the full output contract — then adjust the schema, format, and strictness.

Open in JSON Output Prompt Builder

Prompt Template

TASK
Summarize the sales call notes in the input into a CRM activity entry.

OUTPUT FORMAT
Return your entire response as a single valid JSON object matching the schema below.

SCHEMA
Each field, its type, and its meaning:
- contact_name (string, required): Person the call was with
- company (string, required): Their company
- call_date (string, required): Date of the call, YYYY-MM-DD
- summary (string, required): Two-sentence call summary
- next_step (string, required): The agreed next action with its date
- deal_stage (string, optional): Pipeline stage after this call

EXAMPLE OF A VALID RESPONSE
{
  "contact_name": "Jane Smith",
  "company": "Acme Corp",
  "call_date": "2026-06-10",
  "summary": "Short free-text description",
  "next_step": "Short free-text description",
  "deal_stage": "standard"
}

VALIDATION RULES
- Every required field must be present.
- Optional fields with no value: set them to null — do not omit the key.
- Field types must match the schema: numbers without quotes, booleans as true/false.

OUTPUT RULES
- Output only valid JSON — nothing else.
- Do not include explanations.
- Do not include commentary.
- Do not wrap the output in markdown.
- Do not use code fences.
- Return exactly one JSON object.
- Follow the schema exactly — every required field present, no extra fields.
- Use null for unknown values — never invent data.
- No trailing commas.

Related Resources

Explore more resources that work well with this one.