Fixing Format and Structure Failures

Format Failures Are the Easiest Prompt Problems to Fix. They Are Also the Most Preventable.

Of all the failure modes in this track, format failures are the most straightforward: the output was in the wrong structure, the wrong length, or the wrong medium. Bullet points when you needed prose. No headers when you needed them to navigate. A 500-word essay when you needed a 3-sentence summary. A numbered list when you needed flowing analysis.

These failures are easy to fix because the cause is almost always the same: you did not specify the format you wanted, so the model applied its default for the task type. Add the format specification, get the format you need. The reason this lesson exists is that format specification is also the most commonly omitted element in intermediate prompting — and when it is omitted, the model's defaults are strong enough to significantly change what the output is useful for.

What This Failure Looks Like

Format and structure failures show up in consistent patterns based on task type:

• Bullet overuse: The model's default for most information tasks is bullet points. Analysis becomes a bullet list. Explanations become enumerated items. Anything with multiple parts becomes a formatted list, even when prose would communicate better.
• Header overuse: For longer outputs, the model defaults to adding section headers. In a document that needed to read as continuous prose, headers impose a structural fragmentation that changes how the content reads.
• Length default: Without a length instruction, models aim for what seems thorough. "Thorough" is calibrated to the task type and produces very different results — a summary without a length constraint might come back at 400 words when you needed 80.
• Wrong medium: The model defaults to prose for explanations and to structured output for comparisons and lists. When the task type implies one format but you need another, the default will not match your need.

What Caused It

Format failures have a single root cause: the output specification was missing from the prompt. Without an explicit format instruction, the model applies the format that is most statistically associated with the task type you described. These defaults are not wrong in a general sense — they are reasonable choices for the average version of that task. They are wrong for your specific version of it when your needs differ from the average.

A secondary cause is worth noting: format failure sometimes signals a task interpretation problem. If you asked for analysis and got a bullet list, the model may have read your prompt as a list-extraction task rather than an analytical one. In that case, the format symptom points to an underlying ambiguity in the task statement — fixing the format instruction alone may not be enough if the model is solving the wrong problem.

The Fix

The primary fix is explicit output specification. Tell the model exactly what the format should be: prose or bullets, headers or not, word count or paragraph count, and what the output should look like when complete.

No format specification: "Compare the pros and cons of asynchronous versus synchronous remote work communication."

With format specification: "Compare the pros and cons of asynchronous versus synchronous remote work communication. Format: two flowing paragraphs — one for async, one for sync. No bullet points, no headers. Under 200 words. End with a one-sentence recommendation for a team that spans three time zones."

When a positive format instruction is not enough — when the model's default is strong enough to override a positive instruction — add a negative constraint:

Positive instruction (sometimes insufficient): "Write this as flowing prose."

Negative constraint added: "Write this as flowing prose. Do not use bullet points. Do not use section headers. Do not add a summary section at the end."

Negative format constraints function as hard limits. "Do not use bullet points" is unambiguous — it either holds or it does not. A positive instruction like "use prose" leaves room for the model to use prose in some sections while reverting to bullets in others. The negative constraint closes that gap.

When This Fix Doesn't Work

Format specification fixes format failures in the large majority of cases. Three situations where additional work is needed:

When the format conflict reflects a task mismatch: If the model keeps producing bullet-point analysis despite explicit prose instructions, it may be because the task description reads as a list task. Review the task statement: does it say "explain" and "analyze," or does it say "list," "identify," and "name"? If the latter, the task description is generating the format, not just the format instruction.

When the format needs vary by section: Some outputs need different structures in different sections — an executive summary in prose, a comparison table for the data, a recommendation in bullet points. Single-pass format instructions cannot handle this. Break the output into sections and specify format for each section separately.

When the correct format is something the model handles awkwardly: Some structured outputs — complex tables, specific document schemas, precise JSON structures — may require template-based prompting or a format example rather than a format description. Providing a sample of what the output should look like is more reliable than describing the format in words when the structure is complex.