Why Default Formatting Is Usually Wrong for You
AI models default to a narrative prose style with some occasional bullet points — a format that emerged from training data that was mostly articles, essays, and documentation. That format works for some use cases and fails for many others. If you're putting output into a spreadsheet, you need a table. If you're feeding it into code, you need JSON or a specific data structure. If you're using it in a chat, you need short lines without markdown syntax. If you're putting it in a slide deck, you need bullet points with no more than 8 words each. Specifying the format in your prompt is how you get output that fits your destination without any reformatting.
Common Formats and When to Use Them
Markdown bullet points work well for content that will be rendered in Notion, GitHub, or a CMS. Plain numbered steps work well for instructions or processes where order matters. Tables are ideal for comparisons, feature matrices, or any structured data with clear categories. JSON is essential when feeding output directly into code or automations. Code blocks with syntax specification (Python, TypeScript, SQL) ensure the model formats code correctly with appropriate language conventions. Plain paragraphs with no markdown are best for email drafts, Word documents, or anywhere markdown characters would render as literal symbols rather than formatting.
How to Specify Format Clearly
The most reliable way to specify format is to describe the exact structure you want, not just the label. 'Bullet points' is less precise than 'each insight on its own line starting with a dash, no sub-bullets, max 15 words per bullet.' 'A table' is less precise than 'a markdown table with 3 columns: Feature, Free Plan, Pro Plan — one row per feature.' The more your format instruction mirrors the exact structure you need, the less cleanup you'll have to do. If in doubt, show an example: paste a 2-row version of the table you want and ask the model to continue in the same format.
Format Instructions for Structured Data Output
When you need AI output to feed into a system or application, format becomes critical rather than optional. For JSON output, specify the exact schema: 'respond only with a JSON object in this format: {"title": string, "tags": string[], "summary": string}.' For CSV, specify delimiter and column headers. For code, specify the language, whether to include comments, and whether to include test cases. The model is quite good at following format specifications for structured data — but it needs an explicit schema, not just a label like 'JSON' or 'CSV.'
Where to Put Format Instructions in Your Prompt
Format instructions work best at the end of your prompt, after role, context, and task. This positioning ensures they apply to the complete response rather than creating confusion about what they modify. The exception is when the format is a core part of the task itself — for example, 'write a table comparing X and Y' where the table instruction is part of the task definition. In that case, the format is naturally embedded in the task. For all other cases, end your prompt with a 'Format:' or 'Output:' section that specifies exactly what you want.
Using Examples as Format Templates
One of the most powerful format techniques is providing a filled-in example and asking the model to follow the same pattern. Instead of trying to describe a complex structure in prose, simply show two or three rows of what you want and end with 'continue this format for the remaining items.' This approach works for tables, reports, analyses, code documentation, and almost any structured output. The model is remarkably good at pattern-matching from examples — often better than it is at interpreting verbal format descriptions, especially for complex or unusual structures.