Controlling Output Format

Why Format Control Matters

The same information can be useless or immediately actionable depending on how it's formatted. An answer that buries the key decision in a wall of prose is far less useful than one that surfaces it in a clean JSON object ready to be parsed by your code, or a bullet list your team can scan in ten seconds. Format instructions are one of the most reliable things you can include in a prompt — models follow them with high consistency when stated clearly.

Markdown Formatting

For responses that will be read by humans in a markdown-rendering interface, you can request specific structure: "Use ## headers for each section", "bold the key terms", "include a horizontal rule between each item", "format any code using triple-backtick code fences with the language specified." For plain text interfaces, avoid requesting markdown — it will render as asterisks and hashes rather than formatted text.

JSON and Structured Schemas

When integrating AI output into code, JSON is almost always the right choice. Specify the exact schema: key names, value types, whether arrays are allowed, and any constraints on values. The more specific your schema instruction, the more reliably the model conforms. For critical production use, consider pairing this with a library like Instructor (Python) or a structured output mode (OpenAI's response_format={"type": "json_object"}) to enforce schema compliance at the API level.

Controlling Length

Length instructions are taken seriously by models when stated explicitly. "Answer in exactly three bullet points," "respond in one sentence," "provide a response between 150 and 200 words" — these constraints shape the output. Without a length constraint, the model fills to what it estimates is appropriate, which is often too long. For most tasks, shorter is better: instruct the model to be concise and it will be.