Structured Output Prompts for Channel Analysis

A structured output prompt tells the model to answer in a fixed shape, usually JSON or a labeled list, so each response lands in the same fields. That shape is what lets you compare one video to the next instead of reading a fresh essay every time.

Loosen the format and the model invents a new structure with every run. It drifts. You spend the evening reformatting instead of reading, and comparison becomes impossible.

Why a fixed shape matters

Free text hides inconsistency. Two answers can sound identical and mean different things. A schema forces the model to commit: this is the hook type, this is the drop-off minute, this is the suggested fix. Commitment is what makes the claim testable.

Because the fields never move, you can sort, count, and chart the output. Analysis is comparison, and comparison needs the same columns on every row.

A schema that covers the basics

Keep the field names plain and stable across runs. Renaming a field between prompts silently breaks every comparison you built on top of it.

• hook_type: a short label such as question, bold claim, or cold open.
• retention_risk: the minute where viewers most likely left.
• evidence: the specific row or transcript line behind the claim.
• confidence: low, medium, or high, never left blank.
• next_test: one concrete change to try on the next upload.

The traps to avoid

The first trap is the open field. A field labeled 'notes' invites the model to ramble, and rambling text is exactly what you were trying to escape when you moved to a schema.

The second trap is a schema with too many fields. Ten required values push the model to fill blanks with guesses. Six fields it can answer well beat twelve it has to invent.

A schema is a promise about where each answer will live. Break the promise and you are back to reading prose.

Making the model stick to it

Repeat the schema in the prompt and ask the model to return nothing else. If your tooling supports it, request valid JSON and validate the response before you trust a single field.

When a field comes back empty or vague, treat that as a signal. Either the data did not support the field or the field was poorly defined. Fix the prompt, not the output by hand.

Key takeaways

Start with four fields, run the prompt on ten videos, then add a field only when you find yourself wanting to sort by it.

• A fixed schema turns analysis into rows you can compare.
• Keep field names stable, since renaming breaks your comparisons.
• Avoid open fields and oversized schemas that invite guesses.
• Validate the output and repair the prompt when a field keeps failing.