Most people spec the topic. Almost nobody specs the format. Then they wonder why the output doesn’t fit where they were going to put it.
Here’s what’s actually happening: when you don’t tell a model how to structure a response, it falls back on the format that most commonly appeared alongside similar content in training — which for most topics is a medium-length blog-post-shaped thing with a few headers and some bullet points. That’s not stupidity. It’s a reasonable prior. But it’s almost never right for the specific thing you’re trying to do.
The format is part of the task, not decoration
A summary you’re pasting into a Slack message needs to be two sentences. A summary for speaker notes needs complete sentences but short ones. A summary for an exec who reads everything on mobile needs a single-sentence verdict first, details after. These are three different outputs for three different situations — and “write me a summary” produces the same shape every time, regardless of which one you actually need.
The default format also changes the content, not just the presentation. Ask for a bulleted list and the model will break things into discrete, roughly-parallel points whether or not the underlying ideas are actually parallel. Ask for prose and it’ll connect things with transitions that sometimes imply a relationship that isn’t there. The format shapes what gets said. If you’re not choosing the format, the model is, and its choice is optimized for “looks complete” rather than “works for what you need.”
What the default actually looks like
The classic AI default: a short intro paragraph, three to five bullet points, a closing sentence. You’ve seen it a thousand times. It’s the format that performs well across the widest range of vague requests, which makes it a statistically rational default and a practically annoying one. Too long for a message, too short for a document, too listy for flowing writing, too vague for anything where sequence or priority matters.
The second most common default is the numbered list with a sentence or two per item. Also fine. Also usually not what you needed.
Neither of these is wrong. They’re just not chosen for your situation — they’re chosen to look like a complete and reasonable response to an underspecified prompt.
Being specific about format is faster than reformatting the output
The fix requires actually deciding what you want before you ask. A few patterns that consistently produce more usable output:
- State the container first. “A three-sentence summary I can paste into an email” front-loads the constraint that matters most. The model builds content to fit that shape rather than writing content and trimming it after.
- Specify length in units that mean something. “Short” is relative. “Under 80 words” or “two paragraphs” is not. Token counts work too if you’re building something programmatic.
- Say whether you want prose or structure, and why. “As a numbered list — I’m presenting this verbally so I need discrete points, not paragraphs” gives the model enough context to make sensible choices within the constraint, not just follow it mechanically.
- Tell it what you’re going to do with the output. This sounds like extra work but it’s the fastest way to avoid the generic default. “I’m dropping this into a project brief” produces a different shape than “I’m texting this to a client.”
Where this matters most: automation
For one-off tasks, a mismatched format is just annoying — you fix it manually and move on. For anything running on a schedule without a human reviewing each output, format drift becomes a real problem.
Running a blog pipeline means every post has to come back in a consistent structure: a specific JSON shape with fields for title, body, tags. If the model decides to respond in a slightly different format — markdown instead of JSON, or a field named differently — the downstream parsing breaks, and depending on how error handling is written, it might break silently. Specifying format isn’t a nicety in that context; it’s the entire thing the prompt is doing. The topic is almost secondary.
The same logic applies to anything that stitches AI output into a larger workflow: meeting summaries, draft replies, report sections. Whatever goes downstream from the model is depending on getting something predictably shaped. A vague format instruction is a reliability risk, not just an aesthetic one.
One caveat worth naming
Tight format constraints can make outputs worse when the task is exploratory. If you’re using AI to brainstorm or work through something ambiguous, forcing a specific structure early constrains the thinking too soon. The five-bullet default is annoying when you need prose, but prose is annoying when you needed quick options to scan.
It’s worth knowing which mode you’re in before you start: exploring, or producing? Exploratory prompts benefit from loose format constraints. Production prompts need tight ones. Most format frustration comes from blurring those two, not from the model making bad choices.
If you’re looking at your last ten prompts and none of them specified a format, pick one you care about and try adding the format spec before anything else — before the topic, before the context. See what changes. It usually does.
