Why is this the cheapest improvement available?

Because it costs one sentence and removes the reformatting that most people were doing by hand. The content was already right in most disappointing cases; the shape was the problem, and it is the part a request can fix immediately without any other change.

When does asking stop being enough?

When something other than a person reads the answer. A reader notices an unexpected shape and adapts; a downstream step either breaks or misreads it and continues. At that point the shape is an interface and needs a mechanism rather than a request.

Does constraining the shape make answers more accurate?

No, and expecting it to is a common disappointment. It removes malformed output entirely and touches nothing else, so a system can go from occasionally unparseable to reliably well-formed and wrong. Those are separate problems with separate fixes, and the second one is not addressed by anything in the request. A check on the content has to sit somewhere else entirely.

Glossary

Output format

The output format is the shape an answer is asked to arrive in, such as a table or a list, and asking for one is the cheapest change that makes results usable.

In plain terms

Saying how you want the answer to come back. A table with these columns, a list of five, a short paragraph. It costs nothing to ask and it is the difference between something you have to reformat by hand and something you can use straight away.

Why it matters

Because most disappointment with an otherwise good answer is about shape rather than content. The material was right and it arrived as three paragraphs when what was needed was a table, and reformatting that by hand is the work the tool was supposed to remove.

How it works

Asking is usually enough when a person is reading the result. Say what shape is wanted and it generally arrives that way, and where it does not the reader can see immediately that it has not. That feedback loop is what makes the informal version work: the failure is obvious and cheap.

It stops being enough the moment something automated depends on it. A person seeing an unexpected shape adjusts; a downstream step receiving one either fails or, worse, carries on with something it has misread. At that point the shape has become an interface, and asking politely for an interface is not how interfaces are held.

Products draw this distinction in their own settings, which is the clearest signal that it is real. Modes that constrain the shape during generation are sold as reliability features rather than as prompting advice, and their existence is an admission that a request alone is not a guarantee.

The shape and the content are separate properties and can fail independently. A perfectly formed table can contain the wrong figures, and a correct answer can arrive in the wrong shape. Constraining the format fixes exactly one of those, which is worth remembering when a system stops producing malformed output and starts producing well-formed nonsense.

Who is reading the answer

This distinction is easy to miss because the transition happens without anybody deciding it. A request is written while a person is reading the answers, asking for a shape works, and everybody is satisfied. Later the same request is wired into something: the output goes to a spreadsheet, or feeds a following step, or lands in a system that expects columns. Nothing about the request changed and its guarantees did not improve, but the consequences of an unexpected shape have moved from a reader's mild irritation to a downstream failure nobody is watching for. The signal that this has happened is rarely dramatic. It is simply that the answer has stopped being read by a human being, and from that moment a politely worded request is holding up something that needs a mechanism.

Seen in the wild

An assistant asked for a table with named columns rather than prose, where the reader sees immediately whether the shape arrived.
ChatGPT
An automation step whose result feeds the next step, where an unexpected shape is not noticed by anybody before it is acted on.
Make
A self-hosted platform where visual nodes accept real code when the logic demands it, which is where a shape stops being a request and becomes a contract.
n8n

Common misconceptions

People assume

Asking for a shape guarantees it.

In fact

It requests it, and requests are followed rather than enforced. That is entirely adequate while a person is reading the result and inadequate the moment something automated depends on it, which is why products sell separate settings that constrain the shape while it is being produced.

People assume

A well-formed answer is a correct one.

In fact

Shape and content fail independently. A table with the right columns and the wrong numbers passes every structural check there is. Constraining the format removes malformed output and leaves every other kind of wrong exactly where it was, which is a smaller improvement than it feels like.

Telling them apart

Output format vs Structured output

Output format

Asking for a shape, in the request.

Structured output

A product setting that constrains the shape as it is produced.

One is a request and the other is a guarantee. Which you need depends on whether a person or a system reads the result.

Questions

Why is this the cheapest improvement available?: Because it costs one sentence and removes the reformatting that most people were doing by hand. The content was already right in most disappointing cases; the shape was the problem, and it is the part a request can fix immediately without any other change.
When does asking stop being enough?: When something other than a person reads the answer. A reader notices an unexpected shape and adapts; a downstream step either breaks or misreads it and continues. At that point the shape is an interface and needs a mechanism rather than a request.
Does constraining the shape make answers more accurate?: No, and expecting it to is a common disappointment. It removes malformed output entirely and touches nothing else, so a system can go from occasionally unparseable to reliably well-formed and wrong. Those are separate problems with separate fixes, and the second one is not addressed by anything in the request. A check on the content has to sit somewhere else entirely.

Key takeaways

Asking is enough while a person reads the result, and only until then.
Once something automated depends on the shape, it has become an interface.
Product settings that constrain the shape exist because a request is not a guarantee.
Shape and content fail independently; fixing one leaves the other untouched.

Tools that use this

ChatGPT
A table with named columns, where the reader sees the shape immediately.
Make
A step whose result feeds the next, with nobody checking the shape.
n8n
Nodes accepting real code, where a shape becomes a contract.