Thinkorithm

A Practical Guide to Working with AI Assistants

Simple habits that make AI assistants more useful and less risky, drawn from published public-sector guidance.

Published · 2 min read

AI assistants can save time on writing, summarising and research, but only if they are used with some care. The habits below apply to any assistant, regardless of which product or version you use.

1. Decide what the tool is for

Start with the task, not the tool. Assistants are generally good at drafting, rewording, summarising and generating ideas. They are less suitable when an answer must be exactly right and cannot easily be checked, or when the decision has serious consequences for someone.

2. Give context and a clear instruction

An assistant knows nothing about your situation beyond what you tell it. State who the result is for, what it should contain, what to avoid and what form it should take. Short examples of the output you want are often more effective than long descriptions. Ready-made patterns are collected in the prompt library.

3. Protect information

Treat anything you type into an online service as leaving your control unless you know otherwise. Do not paste passwords, personal data about other people, or confidential material into tools that are not approved for it. Check the service’s data and privacy settings before relying on it.

4. Check facts independently

Language models can produce confident but false statements, including invented references. Verify names, figures, dates, quotations and citations against an original source before you use them. If an assistant cites a source, open it and confirm it says what is claimed.

5. Keep a person responsible

An assistant can propose; a person should decide. This matters most when the output affects other people — in hiring, health, finance, education or legal matters. Be open about where AI has been used when that is relevant to the reader.

6. Iterate

A first answer is a draft. Ask for corrections, a different structure, a shorter version or a list of assumptions the assistant made. Asking the assistant to point out weaknesses in its own answer often surfaces problems worth checking.

A short checklist

  • Is this a task the tool is suited to?
  • Did I give enough context and a clear instruction?
  • Did I keep sensitive information out?
  • Have I checked every fact I intend to rely on?
  • Is a person accountable for the final result?

Sources

  1. UK Government. Artificial Intelligence Playbook for the UK Government, 2025.
  2. National Institute of Standards and Technology. Generative Artificial Intelligence Profile, NIST AI 600-1, 2024.
  3. Brown, T. B. et al. Language Models are Few-Shot Learners. NeurIPS 2020.