Lesson 05: Prompt Engineering for Research¶
Prompt engineering is the practice of writing inputs to an LLM that reliably produce useful outputs. It is not a mystical art. It is the application of clear communication to a system that responds well to clear communication.
This lesson covers the techniques that matter most for research tasks: summarisation, explanation, brainstorming, and structured output.
The mental model: briefing a smart assistant¶
Write your prompt as if you were briefing a knowledgeable colleague who has just joined the project. They are capable but know nothing about your specific context unless you tell them. They will do exactly what you ask, not more, not less, so be specific about what you want.
Core techniques¶
1. Give context before the task¶
Bad:
"Summarise this."
Better:
"I am a PhD student in astroparticle physics writing a thesis introduction. Summarise the following abstract in two sentences suitable for a specialist audience."
The context shapes the register, the level of detail, and what counts as important.
2. Specify the format of the output¶
If you want a bullet list, say so. If you want a single paragraph, say so. If you want structured output (JSON, a table, a numbered list), describe the exact format.
"Return your answer as a markdown table with columns: Term | Definition | Example." "List the five most important points as a numbered list. One sentence per point."
3. Set the audience and level¶
"Explain this to a physicist who has not worked in machine learning." "Write this for a general audience with no physics background." "Assume I am familiar with the standard model but new to cosmology."
The model will adjust vocabulary, assumed knowledge, and depth accordingly.
4. Ask for reasoning (chain-of-thought)¶
For complex problems, asking the model to reason step by step before giving an answer often produces more accurate results.
"Think through this step by step before giving your answer." "First identify what information is needed, then answer the question."
This is especially useful for checking whether the model's reasoning is sound, even if the answer is ultimately wrong.
5. Give examples (few-shot prompting)¶
If you want output in a specific style or format, the most reliable way to get it is to show an example.
"Rewrite the following sentences in the style of the example below.
Example input: The measurements were performed at low temperature. Example output: Measurements were performed at low temperature.
Input: The analysis was conducted using the software developed by the team."
6. Constrain what to avoid¶
"Do not include citations." "Do not use bullet points." "Do not suggest anything that requires internet access."
Constraints prevent the model from producing output you will have to discard.
Prompts for common research tasks¶
Summarisation¶
"Summarise the following [abstract / paper / section] in [N sentences / one paragraph / a bullet list]. Audience: [graduate students / specialists / general public]. Focus on: [the main result / the methodology / the limitations]."
Concept explanation¶
"Explain [concept] at the level of a [first-year PhD student / undergraduate / expert]. Cover: what it is, why it matters, and the main caveats. Do not use analogies I would need to fact-check."
Brainstorming¶
"I am investigating [problem]. Generate [N] hypotheses or approaches I might not have considered. For each, give one sentence on why it might be worth pursuing and one sentence on its main weakness."
Code generation¶
(See Lesson 02 for code-specific prompts.)
Structured extraction¶
"From the following text, extract all named physical quantities, their values, and their units. Return a markdown table with columns: Quantity | Value | Unit | Source sentence."
Argument strengthening¶
"I am making the following argument in a paper: [argument]. Identify the two weakest points in this argument and suggest how to address each."
System prompts and persistent context¶
If you are using an LLM via an API or a tool that supports system prompts, you can provide persistent context that applies to all messages in a session. This is useful for:
- Setting your role and background once rather than repeating it every prompt.
- Enforcing output format constraints.
- Specifying style conventions.
Example system prompt for research writing:
You are assisting a PhD student in astroparticle physics.
The student is fluent in physics but not in machine learning.
Respond concisely. Use LaTeX for all mathematical expressions.
Do not generate citations — the student will find references independently.
Flag any claim you are uncertain about.
Temperature and parameters¶
Most LLM interfaces expose at least one generation parameter:
Temperature controls randomness. A higher temperature (e.g. 0.9) makes the model sample more diverse tokens; a lower temperature (e.g. 0.2) makes it more deterministic and focused.
| Task | Temperature guidance |
|---|---|
| Factual extraction or summarisation | Low (0.0–0.3): you want the most likely correct answer |
| Brainstorming or creative drafting | Higher (0.7–1.0): you want variety |
| Code generation | Low to medium (0.0–0.5): you want correct, predictable output |
For most research tasks, the default temperature (typically 0.7–1.0) is fine. If you are using the API and need reproducible output, set temperature to 0 and fix the random seed if the API supports it.
Iterating on prompts¶
Your first prompt is rarely optimal. Prompt engineering is an iterative process:
- Write a first prompt and look at the output.
- Identify what is wrong: too long, wrong level, missing something, wrong format.
- Revise the prompt to address that specific issue.
- Repeat until the output is useful.
Keep a record of prompts that work well. A prompt that reliably produces good summaries of astrophysics abstracts is worth saving.
Industrial context: prompt engineering as a discipline¶
Prompt versioning. Organisations that deploy LLMs in production treat prompts as code: they are version-controlled, reviewed, and tested. For research workflows you use repeatedly, document the prompts that work.
Prompt injection. When user-supplied text is incorporated into a prompt, adversarial inputs can manipulate the model's behaviour. If you are building an automated pipeline that feeds external data to an LLM, sanitise inputs and constrain the model's output format.
Structured outputs. Modern APIs support JSON mode or structured output schemas, where the model is constrained to produce valid JSON matching a specified structure. This is much more reliable than asking the model to "return a JSON object" in free text. Use structured outputs for any pipeline that processes LLM output programmatically.
Evaluation. Industrial teams evaluate prompts systematically: a prompt change is tested on a benchmark set of inputs before deployment. For research, this means testing a new prompt on several representative examples before relying on it for important work.
Checklist¶
- [ ] My prompts include relevant context (my role, the purpose of the task, the audience).
- [ ] My prompts specify the desired output format.
- [ ] I iterate on prompts rather than accepting the first output.
- [ ] I have saved prompts that reliably produce useful results.
- [ ] I use low temperature for tasks where accuracy matters.
- [ ] For automated pipelines, I use structured output formats.
What to read next¶
Lesson 06 brings everything together in a verification checklist: the habits that separate researchers who use LLMs safely from those who do not.