structure-sections — Prompt Organization

Severity: WARN · Auto-fix: Yes · Category: 🏗️ Structure

What It Does

Checks that the prompt has some form of explicit structure. Structureless prompts — a wall of text with no sections — produce inconsistent output and are harder to maintain and iterate on.

What Counts as "Structured"

The rule considers a prompt structured if it contains any of these:

Signal	Example
XML tags	<task>, <context>, <output_format>, <role>
Labeled headings	Task:, Context:, Output:, Goal:, Requirements:, Instructions:
Markdown headers	## Task, # Instructions, ### Output Format
JSON structure	{"key": "value"} or ["item"] with quotes
Code fences or ---	```python or --- horizontal rules
Numbered sections	1. , 2. , 3. at the start of lines

The bar is intentionally low

If your prompt has any structure at all — even just 1. Do X\n2. Do Y — this rule won't fire. It's looking for completely unstructured walls of text, not enforcing a specific structure format.

When the Rule Fires

Fires:

Write a function that parses JSON data and returns a dictionary handling
errors gracefully and logging exceptions and returning None on failure.

No tags, headings, numbering, or delimiters → WARN.

Does NOT fire:

1. Parse JSON data from the input string.
2. Return a dictionary on success.
3. Return None and log the exception on failure.

Numbered list = structured → no finding.

Example Output

[ WARN ] structure-sections (line -)
  No explicit sections detected (Task/Context/Output).

[ INFO ] structure-recommendations (line -)
  Recommended templates: headings (Task:, Context:, Output:) /
  XML tags (<task>) / Markdown (## sections)

Two findings are emitted: the WARN for the missing structure, plus an INFO recommendation suggesting format options.

Auto-Fix

When --fix is passed, PromptLint wraps the prompt in a <task> tag (or uses the configured structure_style):

Before:

Write a function to parse JSON.

After (structure_style: xml):

<task>
Write a function to parse JSON.
</task>

Auto-fix only adds a task wrapper

The scaffold only wraps the existing content in a <task> tag. It does not infer and add <context> or <output_format> sections — those require understanding your intent. Use the scaffold as a starting point and add the other sections manually.

Configuration

structure_style: xml   # "xml" | "headings" | "markdown" | "auto" | "none"

rules:
  structure_sections: true

fix:
  structure_scaffold: true

structure_style	Auto-fix output
xml (default)	<task>...</task>
headings	Task:\n...
markdown	## Task\n...
auto	Detects existing style and matches it
none	Auto-fix disabled for this rule

False Positives

Short prompts — a 10-word prompt like "Translate this to French" is fine unstructured. The rule fires regardless of prompt length. If you have intentionally terse prompts (translation, simple Q&A), disable the rule or use a short labeled heading.

Prompts that are themselves JSON/code — if your prompt is a JSON object or a code snippet, the rule won't fire (those count as structured). But if you're asking about code and the prompt contains a code fence, it also won't fire.

Why Structure Matters

Structured prompts:

• Give the model explicit signal about what's task vs. context vs. format
• Are easier to maintain as your requirements evolve
• Enable reliable few-shot patterns when the structure is consistent
• Make it obvious when sections are missing (e.g., <output_format> not present)