[Disclosure Agreement 2] Combining Schema, Prompt and Validation: Reduce AI illusion and error write-back

Direct answer: Schema cannot allow AI to completely eliminate hallucinations, but it can significantly reduce the chance of errors entering the system. AI Ready should integrate prompt, data schema, output schema, field whitelist, verification failure handling and human review into a set of processes.

Who should read this?#

• AI prompt word engineer and system integration planner.

• Backend engineers who need to make model output writeable back to JSON.

• The ecommerce technical team that is processing the automation of products, orders, and customer support data.

Why is Prompt alone not enough?#

Prompt can tell the model "please return JSON" or "don't make up product information", but the model may still:

• Return non-existent fields.

• Correct price, inventory or material errors.

• Produces unparseable JSON.

• Making unauthorized promises in customer support responses. -Write the recommendations into a formal decision.

Therefore, AI Ready requires prompt followed by validation. The model output is not the end of the process, but data to be verified.

Schema Context: Let the model understand the available fields#

Before sending the task, the system should provide the necessary schema context:

{
  "resource": "product",
  "writable_fields": {
    "draft_short_description": "string, max 280 chars",
    "draft_long_description": "string, max 3000 chars",
    "meta_title": "string, max 70 chars",
    "meta_description": "string, max 155 chars"
  },
  "read_only_fields": ["sku", "price", "stock_quantity", "certifications"],
  "required_output_fields": ["draft_short_description", "meta_description"]
}

This lets the model know which fields can be written to and which can only be referenced.

Output Schema: Allow the system to verify#

{
  "type": "object",
  "required": ["draft_short_description", "meta_description"],
  "additionalProperties": false,
  "properties": {
    "draft_short_description": {
      "type": "string",
      "maxLength": 280
    },
    "meta_description": {
      "type": "string",
      "maxLength": 155
    },
    "faq": {
      "type": "array",
      "maxItems": 5
    }
  }
}
`

additionalProperties: false` is important because it prevents the model from adding fields on its own.

What to do when verification fails?#

Don't automatically fix all errors. Suggested hierarchical processing:

• JSON cannot be parsed: The model is required to try again. If it still fails, it will be handled manually.

• If the field length exceeds: you can request the model to be shortened.

• Forbidden field appears: reject the output and log the event.

• The product facts are inconsistent: human review will be performed.

• High-risk commitments: intercept and flag risks.

Prompt design example#

你是電商商品內容助理。
只能根據 data.attributes 與 data.product_name 產生內容。
不得新增價格、保固、認證、醫療效果或庫存資訊。
輸出必須符合 output_schema。
如果資料不足，請在 notes 欄位列出需要人工補充的資訊。

The goal of Prompt is not to let the model "play freely", but to allow it to complete specific tasks within the boundaries of the system.

Manual review cannot be omitted.#

Even if the schema verification passes, it does not mean that the content is necessarily correct. Schema can only verify the format, not the business facts. The following content is recommended for human review:

• High traffic product pages.

• Regulation-sensitive categories.

• Prices, discounts, refunds.

• Medical, food, and safety related descriptions.

• Multilingual high-value market.

FAQ#

Can Schema completely eliminate AI illusions?#

cannot. Schema can limit formats and fields to reduce the chance of errors entering the system, but it still requires data verification, fact checking, and human review.

Is Prompt Engineering still important?#

Important, but it should be used in conjunction with schema, validation, permissions and auditing. Relying on prompt alone to control system risks is not enough.

Under what circumstances can automatic write-back be performed?#

Only low-risk, rollbackable, field-restricted, and validated content is suitable for automatic write-back, such as internal drafts or image ALT candidates. The official page is still recommended for review.

References#

• JSON Schema, https://json-schema.org/
• Google Search Central: structured data introduction, https://developers.google.com/search/docs/appearance/structured-data/intro-structured-data