Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
Yes—an AI agent can turn a natural-language task description into reliable JSON, but only when you define the record first, constrain the output, and validate the result before using it. A schema makes fields and types predictable; it does not prove that the model understood every sentence or avoided an unsupported assumption. The dependable pattern is: design a schema, extract only grounded facts, generate with structured-output support, validate in your application, apply domain checks, and measure errors on representative examples.
1. Define the record before you prompt
Start with the data contract your application needs, not with a vague instruction such as “summarize this task.” Decide the field names, data types, required versus optional values, allowed enumerations, and how missing information is represented. Add examples for terms that could be interpreted in more than one way.
Example schema
The following JSON Schema describes a task record. It deliberately keeps uncertainty explicit instead of forcing the agent to guess.
Recommended Free Tools
{
"$schema": "https://json-schema.org/draft/2020-12/schema",
"type": "object",
"additionalProperties": false,
"required": ["title", "priority", "due_date", "assignee", "actions", "unknowns"],
"properties": {
"title": {"type": "string"},
"priority": {"type": "string", "enum": ["low", "medium", "high", "urgent", "unknown"]},
"due_date": {"type": ["string", "null"], "format": "date"},
"assignee": {"type": ["string", "null"]},
"actions": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": false,
"required": ["description", "status"],
"properties": {
"description": {"type": "string"},
"status": {"type": "string", "enum": ["pending", "completed", "blocked"]}
}
}
},
"unknowns": {"type": "array", "items": {"type": "string"}}
}
}In your prompt, define what each field means. For example, “due_date is the date explicitly stated by the user; do not calculate one from a duration unless the input supplies a reference date.” Tell the agent to put absent facts in null or unknowns, according to the contract.
2. Give the agent an extraction task
Provide the original task text as data and separate it from the instructions. Require evidence-based extraction: copy names and identifiers accurately, preserve the user’s dates and units, and distinguish an explicit statement from an inference. A useful instruction is:
Extract a TaskRecord from TASK_TEXT. Use only facts stated in TASK_TEXT. Never invent a person, date, priority, identifier or completed action. If a required fact is absent, use the schema's null/unknown representation and add a concise explanation to unknowns. Return only the schema-defined object.For long descriptions, ask the agent to retain source spans or short evidence strings in a separate field. This lets reviewers trace a value back to the input without treating a model-generated explanation as proof.
3. Use constrained structured generation
When your model API supports a response schema or strict function arguments, pass the schema directly instead of relying on prose instructions alone. OpenAI’s Agents SDK documents output schemas that validate and parse model output. OpenAI’s function-calling documentation describes strict Structured Outputs that match generated arguments to a supplied JSON Schema. Google and Microsoft document comparable schema-based patterns for structured generation in agent workflows, and Snowflake documents structured output in its Cortex Code Agent SDK.
“Strict” does not mean “true.” It means the returned value conforms to the declared shape and permitted types (subject to the provider’s documented schema subset). The model can still select the wrong enum, omit a relevant detail that was not required, or make an inference that is syntactically valid. Keep the source text and the extraction result together so later checks can detect those errors.
Minimal OpenAI-style Python pattern
The exact client method and model name depend on the SDK version you deploy. The important parts are the schema, an explicit extraction instruction, and application-side validation.
import json
from jsonschema import validate, ValidationError
from openai import OpenAI
client = OpenAI()
TASK_SCHEMA = {"type": "object", "additionalProperties": False,
"required": ["title", "priority", "due_date", "assignee", "actions", "unknowns"],
"properties": {
"title": {"type": "string"},
"priority": {"type": "string", "enum": ["low", "medium", "high", "urgent", "unknown"]},
"due_date": {"type": ["string", "null"]},
"assignee": {"type": ["string", "null"]},
"actions": {"type": "array", "items": {"type": "object", "additionalProperties": False,
"required": ["description", "status"], "properties": {
"description": {"type": "string"},
"status": {"type": "string", "enum": ["pending", "completed", "blocked"]}}}},
"unknowns": {"type": "array", "items": {"type": "string"}}
}}
text = "Prepare the Q4 launch checklist for Maya by Friday. Security review is blocked."
response = client.responses.create(
model="YOUR_MODEL",
input=[{"role": "system", "content": "Extract only facts stated in the task. Use unknowns for missing facts."},
{"role": "user", "content": text}],
text={"format": {"type": "json_schema", "name": "task_record", "schema": TASK_SCHEMA, "strict": True}}
)
record = json.loads(response.output_text)
validate(instance=record, schema=TASK_SCHEMA)
print(record)Check the current SDK documentation for the syntax supported by your installed version. If a provider cannot enforce the schema at generation time, still parse the returned JSON and run the same validator before storing or acting on it.
4. Validate structure, then validate meaning
Validation has two layers:
- Schema validation: Parse JSON and reject wrong types, missing required properties, extra properties, malformed dates and values outside enumerations.
- Task-specific validation: Check business rules and grounding against the source text.
Examples of semantic checks include verifying that every action description has supporting words in the input, rejecting a due date earlier than the stated request date, checking identifier formats, and requiring human review when two people or dates could match the same phrase. A schema-valid object should never bypass authorization, accounting, scheduling or other consequential controls.
Free tools Windows power users keep installed
One-click scans. No signup required.
Handle failures as states, not exceptions to hide
- Missing field: keep the record with
nullorunknownwhen allowed, otherwise route it for clarification. - Invalid JSON or schema error: retain the raw response for diagnostics, retry with a bounded policy, then mark the job failed.
- Refusal or incomplete output: surface the provider status and ask for a narrower extraction rather than silently inserting defaults.
- Ambiguity: record competing interpretations or request a user decision.
- Unsupported inference: reject or downgrade the field and include the reason in an audit log.
5. Build an agent workflow around the extractor
An agent may call tools, retrieve documents or ask follow-up questions before returning the final object. Keep tool results separate from user-provided facts, label their provenance, and apply the same schema to the final response. A practical sequence is:
- Receive and normalize the task text.
- Classify whether it contains enough information for extraction.
- Call retrieval or business tools only when the schema permits externally sourced values.
- Generate the schema-defined record.
- Parse and validate it.
- Run grounding and domain checks.
- Persist the record with source text, model/version metadata, validation status and timestamps.
- Ask for clarification or send to human review when checks fail.
Do not let a tool call mutate production data merely because the model produced valid JSON. Gate side effects on authorization, confidence rules you have tested, and (for high-impact actions) explicit approval.
6. Evaluate extraction quality instead of assuming it
Create a small, representative test set from real task descriptions. Label the expected value for every field, including intentionally missing and ambiguous cases. Track these categories separately:
- missing fields that should have been extracted;
- incorrect values or wrong enum choices;
- unsupported inferences;
- schema or parsing failures;
- unresolved ambiguity and unnecessary escalations.
Run every candidate model or platform against the same texts, schema and scoring rules. Documentation establishes available mechanisms, not a provider-neutral winner for this exact workflow. Compare latency, cost, observability, deployment constraints, tool support and failure reporting only after you have measured your own workload.
7. Platform decision checklist
| Question | What to verify |
|---|---|
| Schema enforcement | Supported JSON Schema subset, strict mode, and where validation occurs |
| Parsing | Native SDK parsing, typed objects, and how failures reach your code |
| Agent tools | Whether tools can be used while the final answer remains schema-defined |
| Failure handling | Refusals, incomplete output, invalid values and retries |
| Operations | Latency, cost, logging, data handling and deployment limits for your region and plan |
| Evidence | Results from your labeled test set, not schema conformance alone |
8. Troubleshooting common problems
The model returns prose around the JSON
Use a response-schema or strict function-call mode. If unavailable, extract the JSON portion only as a temporary compatibility measure, then validate it and log the raw response. Do not accept regex-cleaned text as proof of correctness.
Dates are inconsistent
Specify an ISO format, timezone and reference date in the schema instructions. Reject impossible dates and keep an explicit null when the task says “next Friday” but no locale or reference date is available.
Required fields are filled with guesses
Change the schema to allow null or an unknown enum, add a no-inference instruction, and test with deliberately incomplete examples. Treat a filled field without evidence as a semantic failure.
Valid output misses an important sentence
Add representative examples, include an actions array or evidence fields, and score recall by field. If completeness matters, run a second coverage check that compares extracted claims with the source text.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsRetries repeat the same error
Capture the validation error and send a targeted correction request, such as “priority must be one of these values; source text does not state priority, use unknown.” Limit retries and route persistent failures to review.
Or skip the browser setup
If your agent workflow needs screenshots of task-related pages, ScreenshotNeo provides a single API call instead of maintaining browser automation. It removes cookie banners, newsletter popups and chat widgets before capture; bot checks, blank pages and failed loads are never billed; its MCP server lets Claude, Cursor and other MCP clients call take_screenshot, get_page_info and capture_pdf. The Free plan includes 1,000 screenshots a month with no card, and paid plans start at $5 for 3,000.
cURL:
curl -G "https://api.screenshotneo.com/v1/shot" -d access_key=YOUR_API_KEY --data-urlencode url=https://stripe.com -o shot.webpPython:
import requests
r = requests.get("https://api.screenshotneo.com/v1/shot", params={"access_key": "YOUR_API_KEY", "url": "https://stripe.com"}, timeout=90)
open("shot.webp", "wb").write(r.content)Node.js:
const q = new URLSearchParams({ access_key: 'YOUR_API_KEY', url: 'https://stripe.com' });
const res = await fetch(`https://api.screenshotneo.com/v1/shot?${q}`);See the ScreenshotNeo documentation for options and response headers. Sign up free to get 1,000 screenshots each month with no card.
FAQ
Can JSON Schema guarantee factual accuracy?
No. It constrains structure and types. Grounding checks, semantic rules and evaluation are still required.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Should absent values be empty strings?
Usually not. Use null, an explicit unknown enum, or a documented omission so downstream code can distinguish “not provided” from “provided but empty.”
Do I need an agent, or is a single model call enough?
A single structured call is sufficient for straightforward extraction. Add an agent only when retrieval, tools, clarification or multi-step validation materially improves the workflow.
How many examples should I include?
Use examples for ambiguous fields and edge cases, then verify their effect on a labeled test set. More examples are not automatically better if they conflict with the schema.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →
