Practical AI tools for everyday life and business. Explore what AI can do
Learning Hub
PRACTICAL AI FIELD GUIDE

Better prompting: practical recipes

Write a clear request, compare the result and keep the version that works for your task.

Free guide8 min readReviewed

These original examples work as starting points across chat tools. Agent actions, file access and structured-output settings depend on the product. A prompt cannot grant access to a file or enforce a security policy.

Your first useful task

Start with one small win

  1. Write the outcome in one sentence and identify the reader or user.
  2. Add the source, one useful example and any real constraints.
  3. Run a small example, compare the result with your checklist, then revise the weakest part.

Practical tips that make a difference

Use only the context that changes the answer

Include the goal, relevant material, desired output and boundaries when they matter. A specific request is easier to assess than a long instruction to be brilliant. Leave out unrelated history that could distract from the current task.

Based on: OpenAI · Prompting ↗

Demonstrate the difficult part

Use a short input/output example to show the format or judgment you need. Add a case with missing information if that happens in real work. Keep examples consistent; conflicting examples make it harder to know what a correct result should look like.

Based on: Google · Gemini prompt design ↗

Define quality before asking for variations

Choose criteria such as factual accuracy, required sections and a word limit. Compare outputs against those criteria. If the task fails repeatedly, check the missing data or the task design instead of endlessly adding stronger adjectives to the prompt.

Based on: Anthropic · Prompt engineering overview ↗

Improve one thing at a time

Ask for a targeted edit while preserving the parts already accepted. Our recommendation: keep a small set of representative inputs and record which prompt version worked better. That turns personal preference into a repeatable comparison.

Based on: Anthropic · Prompt engineering overview ↗

Developer notes

Turn “works” into acceptance cases

For a coding task, provide representative inputs, expected outputs and failure cases. A generated test should express the required behavior rather than merely repeat the implementation. Then run it in the actual project.

Based on: GitHub · Copilot prompt engineering ↗

Distinguish a response format from a guarantee

“Return JSON” in a chat is an output request. For an API workflow needing a predictable structure, use a supported schema facility and validate the result. You still need to reject incorrect dates, quantities or business rules.

Based on: Google · Structured outputs ↗

Keep reusable rules short

Put stable project commands and conventions in the instruction mechanism your tool supports. Give the individual task its own brief. If a rule protects production or secret data, use actual permission or review controls as well as wording.

Based on: Anthropic · Project memory and instructions ↗

Copy, adapt, then check

Original Utilising AI examples. Replace bracketed placeholders before use. These buttons copy text only; they do not run code or contact an AI provider.

A compact reusable brief

Outcome: [what I need].
Audience: [who will use it].
Source of truth: [attach or paste material].
Required output: [format, length, fields].
Constraints: [budget, time, facts to preserve].
Missing information: identify it; label assumptions.
Check: [how we will judge the result].

Original Utilising AI template. Fill only the fields useful for your task.

Based on: OpenAI · Prompting ↗

Improve an existing draft

Revise the draft below for [audience]. Improve [one issue]. Preserve the accepted facts, structure and call to action. Return the revised draft and a short list of substantive changes. Flag unsupported claims.

Draft:
[paste draft]

Use permitted, non-sensitive source material.

Based on: Microsoft · Effective prompts in Word ↗

Create a testable coding brief

Feature: [behavior].
Existing pattern to follow: [file].
Examples: [input → expected output].
Edge cases: [empty, invalid and boundary input].
Constraints: [interfaces to preserve].
Deliverable: a minimal change, relevant tests, and observed check results.

Original example; the coding agent needs actual project access to run checks.

Based on: GitHub · Copilot prompt engineering ↗

Check your result

  • Is there one main outcome?
  • Does the prompt identify facts and unknowns separately?
  • Can you tell whether the result meets the stated criteria?

Common mistakes to avoid

  • Collecting magic phrases instead of testing the output.
  • Changing five instructions at once and not knowing which helped.
  • Asking an agent to deploy when you only intended to review a draft.

Sources & review notes

Reviewed on 15 November 2026. Tips are paraphrased from the linked documentation; examples and practical recommendations are our own. This is a documentation review, not a hands-on test of every provider account or feature. Interfaces, plans and commands can change.

Independent learning resource by Utilising AI Pty Ltd. No endorsement by the named providers is claimed. Review your chosen tool’s permissions and data handling before supplying business information.

Keep building your AI skills

Create your free assistant prompt Browse all six guides
Better prompting: practical recipes | UtilisingAI