Gemini Structured Output

Gemini Structured Output

Use this skill when Gemini should return validated JSON instead of prose. It covers the modern google-genai SDK path and a salvage parser for older REST-style responses.

When to invoke

• User says: "gemini structured output" / "gemini json schema" / "pydantic to gemini" / "responseSchema"
• Code in the conversation uses: Pydantic models, JSON validation, or Gemini response schemas.

When NOT to invoke

• The user needs streaming structured output.
• The schema requires deep unions, recursive references, or complex OpenAPI features.

Concrete example

User input:

Make Gemini return a validated verdict object for this idea score.

Output:

from pydantic import BaseModel
from google import genai


class Verdict(BaseModel):
    label: str
    score: int
    reasons: list[str]


client = genai.Client()
resp = client.models.generate_content(
    model="gemini-2.5-flash",
    contents="Score this idea 0-100 and label GO/NO-GO: a CLI that lints SKILL.md files.",
    config={"response_mime_type": "application/json", "response_schema": Verdict},
)
verdict = Verdict.model_validate_json(resp.text)
print(verdict.label, verdict.score)

For REST or older SDK paths, use extract_json() from the asset before Pydantic validation.

Pattern to apply

1. Define the expected output as a Pydantic model.
2. Pass the model as response_schema and set response_mime_type to application/json.
3. Validate resp.text back into the model.
4. If the model wraps JSON in markdown/prose, depth-match the first balanced JSON object and validate that.
5. If Gemini rejects the schema, flatten nested models and avoid $ref, deep oneOf, and deep anyOf.

Reference: assets/structured.py.

Source

Distilled from production use across the author's automation projects. v0.1.0. See also: [[gemini-cost-tracker]].