1.5 Structured Output Prompting

When to use it

Downstream use

You need the output in a specific shape: a table for a spreadsheet, a checklist for a protocol, a structured plan you can paste straight into a document.

Consistency across runs

You are running the same prompt multiple times and need results in the same format every time, for example extracting data from a batch of papers.

Choosing the right format

Not every task needs the same structure. Here is a quick guide:

Format	Best for	Avoid when
Table	Comparing items across shared dimensions, structured data, reference material	Cells would need more than two sentences
Numbered list	Step-by-step instructions, ranked items, sequential processes	Items have no inherent order
Bullet points	3-10 discrete items, quick scanning, brainstorming	Items need comparison across attributes
Headings/sections	Multi-part analysis, reports, long-form output	Short, single-topic answers
Template	Standardised outputs you will reuse (protocols, forms, descriptions)	Highly creative or exploratory tasks
Prose	Nuanced analysis, persuasive writing, discussion sections	Data needs to be scanned quickly

A simple rule from Google’s technical writing guide: if you can rearrange the items and the meaning stays the same, use bullets. If order matters, use numbers.

Best practices

Name the format explicitly - “Respond as a markdown table” or “Provide a numbered list” removes ambiguity. Without this, the AI picks whatever format it likes.
Specify the structure - For tables, list the column headers. For sections, name the headings. For lists, say how many items you want. The more structure you define, the more consistent the output.
Show an example when the format is complex - One well-crafted example of the desired output teaches format more reliably than a paragraph of instructions. The AI mimics the pattern.
Keep reasoning and formatting separate - Research shows that strict format constraints can reduce reasoning quality. For complex tasks, ask the AI to think through the problem first, then present its conclusions in the requested format.
State how to handle missing information - Tell the AI what to do when data is unavailable: “If information is not reported, enter NR.” This prevents the AI from guessing or skipping fields silently.

Ready-to-use prompts

Extract structured data from a research paper:

Extract the following information from the attached study
into a markdown table:

| Field | Extracted Data |
|-------|---------------|
| First author and year | |
| Study design | |
| Sample size (N) | |
| Study organism or system | |
| Location | |
| Key methods | |
| Primary outcome measure | |
| Main finding (with effect size and CI if available) | |
| Limitations noted by authors | |

Rules:
- If information is not reported, enter "NR"
- Include exact values with units
- Quote statistical results verbatim (e.g., "p = 0.03")

Compare methodological approaches side by side:

Compare these three approaches for studying pollinator
decline in Dutch agricultural landscapes:
1. Standardised transect walks
2. Pan trapping
3. eDNA metabarcoding

Format as a markdown table with columns:
| Criterion | Transect walks | Pan trapping | eDNA |

Compare across these criteria:
- Taxonomic resolution
- Spatial coverage
- Seasonal constraints
- Equipment and cost
- Expertise required
- Strengths for my research question
- Key limitations

After the table, recommend the most appropriate approach
for a 3-year monitoring programme at 20 sites with limited
taxonomic expertise, and explain your reasoning in 150 words.

Generate a structured statistical analysis plan:

Create a statistical analysis plan for this study:

Research question: [YOUR QUESTION]
Design: [experimental / observational]
Dependent variable: [name, type, how measured]
Independent variable(s): [name(s), type(s)]
Sample size: [N]
Data structure: [nested / repeated measures / independent]

Provide the plan as numbered steps:
1. Data screening (which assumptions to check, which tests)
2. Descriptive statistics (which measures for each variable)
3. Primary analysis (test name, justification)
4. Post-hoc tests (if applicable, with correction method)
5. Effect size (which measure, interpretation benchmarks)
6. Sensitivity analyses
7. R packages to use

For each step, explain what to do if assumptions are
violated.

Generate a structured sampling protocol:

Generate a vegetation sampling protocol for [habitat type]
using [quadrat / transect / point-quarter method].

Structure with these sections:

## 1. Equipment List
| Item | Specification | Quantity |

## 2. Site Selection
Criteria for selecting sampling locations.

## 3. Field Procedure
Numbered steps a trained field technician can follow
independently.

## 4. Data Recording Sheet
A table with columns matching the variables to record.
Include: date, observer ID, site code, [your variables],
weather conditions, notes.

## 5. Quality Assurance
Calibration, duplicate sampling, and data validation checks.

When prose is better

Structured output is not always the right choice. Use prose when:

Nuanced interpretation is needed - Discussion sections, theoretical arguments, and qualitative findings require narrative flow that tables break apart.
You are persuading - Grant significance sections and cover letters need a compelling story, not bullet points.
Ideas are deeply interconnected - When each point builds on the previous one, a list fragments the logic.

A useful hybrid: use structured prompts to extract and organise your raw material, then write the narrative yourself. The AI handles the tedious extraction; you handle the synthesis.

Key takeaway

Sources

Based on materials from Prompt Engineering Guide, Vanderbilt Prompt Pattern Catalog, Tam et al. (2024) - Format Restrictions and LLM Performance, and Anthropic Prompting Best Practices.

Have structured output templates that work well for your research? Share them with RSO so we can add them to the library.