Prompting

Prompting is the primary interface for directing language model behaviour. The quality of the prompt determines the quality of the output.

Be Specific

Vague prompts produce vague responses. State the format, length, audience, and constraints explicitly.

Vague	Specific
“Write a summary”	“Summarise in 3 bullet points for a non-technical audience”
“Improve this code”	“Refactor this Python function for readability; no logic changes”
“Write a blog post”	“Write a 500-word post for senior engineers on X, technical tone, no intro fluff”

The model will fill gaps in your prompt with its default assumptions. If the defaults don’t match what you want, specify explicitly.

Role and Context

Models respond to role framing. Telling the model it is a senior engineer, a sceptical reviewer, or a technical writer shapes the style and depth of the response.

System: You are a senior backend engineer reviewing a pull request.
        Identify correctness issues, performance risks, and missing tests.
        Be direct. No praise unless there is something genuinely good.

User: [paste PR diff]

Context that helps the model:

Audience (who will read the output)
Purpose (what the output will be used for)
Constraints (length, format, what to avoid)
Background (prior decisions, related code, domain context)

Few-Shot Examples

Providing 2–3 examples of the desired input-output format dramatically improves consistency. Especially useful for structured outputs (JSON, tables) and tone matching.

Extract the action items from each note. Return JSON.

Note: "Call Sarah about the Q3 budget by Friday."
Action items: [{"task": "Call Sarah about Q3 budget", "deadline": "Friday", "owner": "user"}]

Note: "Send the design mockup to the client and schedule a review."
Action items: [{"task": "Send design mockup to client", "deadline": null, "owner": "user"},
               {"task": "Schedule design review", "deadline": null, "owner": "user"}]

Note: [your note here]
Action items:

Chain-of-Thought

For reasoning tasks, instruct the model to think step by step before answering. This reduces errors on multi-step problems.

Solve this problem. Think through each step before giving the final answer.

Chain-of-thought helps most on:

Arithmetic and logic puzzles
Multi-condition reasoning (“if A and B but not C…”)
Code debugging and root-cause analysis

Prompt structure for reliability

For production prompts where consistency matters:

[Role / persona]
[Task description]
[Constraints and format]
[Examples if needed]
[Input: {variable}]
[Output:]

Keep the variable input clearly delimited (XML tags, triple backticks, or labelled sections) so the model doesn’t confuse instructions with input data.