--- title: "Checking AI work and building your prompt toolkit" description: "Why checking matters AI assistants are powerful drafting partners, but you are responsible for what you send, publish or decide. A quick, consistent…" url: https://optimizeall.com/learn/prompt-engineering-foundations/checking-ai-work updated: 2026-10-05 --- Prompt Engineering Foundations · Everyday mistakes and good habits · lesson 16 of 16 · 10 min # Checking AI work and building your prompt toolkit ## Why checking matters AI assistants are powerful drafting partners, but you are responsible for what you send, publish or decide. A quick, consistent checking routine lets you move fast *and* stay trustworthy. ## The 4-point check Before using any AI output, run through four questions: 1. **Did it do what I asked?** Right task, audience, length and format? 2. **Is it true?** Check facts, names, numbers, dates, quotes and sources that matter. 3. **Is it appropriate?** Tone, sensitivity, fairness, and anything that could embarrass you, offend someone or break a rule. 4. **Is it mine?** Does it sound like you or your brand? Would you be comfortable putting your name on it? ## How to check facts efficiently - **Ask the AI to flag uncertainty:** "Which statements in your answer should I double-check?" This will not catch everything, but it points you to the riskiest parts. - **Ask for sources, then open them:** if your tool can search the web, ask for links. Open them and confirm they say what the AI claims. Sources can be misquoted or even made up. - **Cross-check key facts** with an official or reliable source: a government site, the company's own website, a recognised publication. - **Use the right tool for numbers:** a calculator or spreadsheet for sums, not your eyes on an AI answer. ## Appropriateness and fairness AI can reflect stereotypes or assumptions found in the text it learned from. When writing about people, such as job adverts, customer personas or performance feedback, read carefully for unfair assumptions about gender, age, nationality, religion or disability. For advertising, remember that rules about honest claims and disclosure of sponsored content still apply to AI-written text. ## Privacy and confidentiality basics - Know your tool's settings: some tools let you control whether your conversations are used to improve their models. - Follow your workplace policy on which AI tools are approved and what data is allowed. - Remove unnecessary personal or confidential information before pasting. - Be careful with other people's content: summarising a public article is different from uploading a client's confidential contract. ## Being open about AI use Norms differ by workplace, school and platform. Some require you to disclose AI assistance; some platforms label AI-generated images or media. When in doubt, check the rules and be honest. It protects your credibility. ## Building your personal prompt toolkit The best prompters are not people who memorise tricks; they are people who save what works. Build a simple toolkit document with sections like: ```text MY PROMPT TOOLKIT 1. Weekly team update Prompt: "Turn these notes into a weekly update for my team of 6..." Notes: works best if I paste notes as bullets. 2. Customer reply (complaint) Prompt: "Draft a reply to this complaint. Apologise sincerely..." Notes: always check the refund amount. 3. Learning a new tool Prompt: "Explain [tool] to me as a beginner, then quiz me..." ``` Update it whenever a prompt works especially well. Many AI tools also let you save custom instructions (for example your role, location and preferred style) so you don't repeat them every time; check your tool's settings. ## Worked example: a quick checking routine A marketing assistant in Jeddah uses AI to draft a product launch email. 1. **Did it do what I asked?** Yes, but it is 250 words instead of 150. She asks for a shorter version. 2. **Is it true?** The draft says "free delivery across the GCC". Their offer is only within Saudi Arabia. She corrects it. 3. **Is it appropriate?** The tone is fine; the discount wording meets the company's promotion rules. 4. **Is it mine?** She changes two phrases to match the brand's usual voice. Total checking time: about five minutes. Without it, a wrong delivery promise would have gone to thousands of customers. ## Checking answers that come with sources Web search and deep research features now attach links to their answers. That is a big improvement, but it is not the same as being correct. Use this routine on any sourced answer you plan to rely on: 1. **Open at least two of the cited sources.** Does the page actually say what the answer claims? 2. **Check the date.** Is the source current enough for the question (prices, rules and schedules change)? 3. **Check the type of source.** An official government or company page outranks a forum post or an old blog. 4. **Check for gaps.** Ask: "What would change this answer? What did you not find?" ## Hands-on: a reusable checking prompt Before you use an important answer, send this in the same conversation: ```text Review your answer above before I use it. 1. List every factual claim (numbers, dates, names, rules) as a bullet. 2. For each, say whether it came from a source I gave you, a web source you cited, or your general knowledge. 3. Mark the 3 claims I should verify first and explain why. Do not rewrite the answer yet. ``` This will not catch everything, but it tells you where to look first. Then verify those claims yourself. ## Privacy settings to check today Take five minutes to open your assistant's settings and look for: - whether your chats may be used to improve the provider's models, and how to opt out if you want to; - memory settings: what is remembered and how to delete it; - how long chats are kept, and how to delete them; - workplace versions (business or enterprise plans) that your organisation may require for work data. Under data protection laws such as the UK GDPR and EU GDPR, Pakistan's evolving data protection framework, the UAE's federal data protection law and Saudi Arabia's PDPL, organisations have duties when they process personal data. You do not need to be a lawyer, but you should follow your organisation's AI policy and avoid pasting other people's personal data into tools that are not approved. ## Measuring your checking habit Track how many errors your checks catch each week. If you never catch anything, either your checks are too light or your tasks are low risk. If you catch many, tighten your prompts (more context, clearer instructions for missing information) so fewer errors appear in the first place. ## Where to go next You now have the foundations: the T-CAFE structure, clarity, context, examples, iteration, formatting, thinking modes, projects and memory, research with sources, files, photos, voice and images, a library of over fifty patterns, and a checking routine. If you want to go deeper, the Optimize All academy offers the Advanced Prompt Engineering course (context engineering, structured outputs, agents and evaluation) and Multimodal and Reasoning Models in Practice (vision, voice, media generation and thinking models). Keep practising on real tasks; that is where the skill grows. ## Video lecture: Checking AI work and building your prompt toolkit Video: [Checking AI work and building your prompt toolkit](https://www.youtube-nocookie.com/embed/JtHDrK1FOKQ) — Why checking matters AI assistants are powerful drafting partners, but you are responsible for what you send, publish or decide. A quick, consistent… 12 chapters · about 7 minutes · captions and full transcript below. 1. Checking AI work 2. Why check? 3. The four-point check 4. Checking facts efficiently 5. Checking sourced answers 6. Appropriateness and fairness 7. Privacy settings to check today 8. Your prompt toolkit 9. A reusable checking prompt 10. Example 1: a retirement speech 11. Example 2: job adverts (illustrative) 12. Recap and next steps ## Lecture transcript ### Checking AI work A marketing assistant in Jeddah almost sent a launch email promising free delivery across the whole Gulf, when the offer only covered Saudi Arabia. Five minutes of checking saved the company thousands of awkward customer conversations. In this lecture you will learn a four-point check for every AI output, how to verify answers that come with sources, the privacy settings worth checking today, and how to build a personal prompt toolkit. By the end, you will move fast and stay trustworthy. ### Why check? Why does checking matter so much? Because you are responsible for what you send, publish or decide, not the AI. A wrong price, a biased phrase or a leaked customer detail carries your name, not the model's. The good news is that checking does not have to be slow. A consistent routine of a few minutes catches most problems, and it lets you use AI confidently rather than nervously. Think of it as the final quality check on a production line: quick, routine and non-negotiable. ### The four-point check Here is the four-point check. One: did it do what I asked? Right task, audience, length and format. Two: is it true? Check facts, names, numbers, dates, quotes and sources that matter. Three: is it appropriate? Tone, sensitivity, fairness, and anything that could embarrass you, offend someone or break a rule. Four: is it mine? Does it sound like you or your brand, and would you put your name on it? Four questions, a few minutes, every time something matters. ### Checking facts efficiently To check facts efficiently, ask the assistant to flag uncertainty. Try: review your answer before I use it; list every factual claim; say whether each came from a source I gave you, a web source you cited, or your general knowledge; and mark the three claims I should verify first. It will not catch everything, but it points you to the riskiest parts. Then use the right tool for numbers, a calculator or spreadsheet, not your eyes on an AI answer. ### Checking sourced answers Answers with citations need their own routine. Open at least two of the cited sources and confirm they actually say what the answer claims. Check the dates, because prices, rules and schedules change. Check the type of source: an official government or company page outranks a forum post or an old blog. And ask what would change the answer, and what it did not find. Sources can be misquoted, outdated, or attached to sentences they do not support. ### Appropriateness and fairness Fairness matters too. AI can reflect stereotypes found in its training text. When writing about people, such as job adverts, customer personas or performance feedback, read carefully for unfair assumptions about gender, age, nationality, religion or disability. And for advertising, the rules on honest claims and disclosure of sponsored content apply to AI-written text exactly as they do to anything else. ### Privacy settings to check today Take five minutes today to open your assistant's settings. Look for whether your chats may be used to improve the provider's models, and how to opt out if you want to. Look at memory, what is remembered, and how to delete it. Check how long chats are kept. And find out whether your organisation requires a business or enterprise plan for work data. Data protection laws such as the UK GDPR, the UAE's federal data protection law and Saudi Arabia's personal data protection law put duties on organisations that handle personal data. Follow your organisation's AI policy. ### Your prompt toolkit The best prompters are not people who memorise tricks. They are people who save what works. Build a simple toolkit document: a name for each prompt, the prompt itself, when to use it, and what to check. For example, weekly team update, customer complaint reply, learning a new tool. Update it whenever a prompt works especially well, and move the stable parts into custom instructions or projects so you do not repeat them. ### A reusable checking prompt Here is a reusable checking prompt you can send before using any important answer. Review your answer above before I use it. First, list every factual claim, numbers, dates, names and rules, as bullets. Second, for each, say whether it came from a source I gave you, a web source you cited, or your general knowledge. Third, mark the three claims I should verify first, and explain why. Do not rewrite the answer yet. It will not catch everything, but it tells you exactly where to look first. ### Example 1: a retirement speech A simple example of the four-point check. You ask for a short speech for your father's retirement. Did it do what I asked? It is four minutes, you wanted two, so ask for a shorter version. Is it true? It says he worked there for twenty-five years, and it was twenty-two, so fix it. Is it appropriate? One joke might embarrass him in front of his manager, so remove it. Is it mine? It says esteemed colleagues, which you would never say, so change it. Four questions, two minutes, and the speech is right in every way that matters. ### Example 2: job adverts (illustrative) Now a business scenario, with illustrative numbers. A recruitment agency in Riyadh uses AI to draft job adverts, about forty a month. Before posting, the team runs the four-point check. One advert asked for young and energetic candidates. That is an age-related phrase that creates discrimination risk, so it becomes motivated and eager to learn. Another listed a salary range that did not match the client's brief. The team asks the assistant to list every factual claim with its origin, and verifies salary, location and benefits against the client form. It adds about five minutes per advert. Across forty adverts, that is a little over three hours a month, a small price for avoiding a discriminatory advert or a salary dispute. Illustrative figures. ### Recap and next steps To recap. Run the four-point check on anything that matters. Ask the assistant to list and rank its claims, open the sources behind sourced answers, and use proper tools for numbers. Watch for bias, follow privacy rules and be open about AI use where required. And build a toolkit of prompts that work. Try this now: create your toolkit document with at least five prompts you will reuse, each with a note on what to check. You now have the foundations. Keep practising on real tasks; that is where the skill grows. ## Key takeaways - Run a 4-point check on every output: did it do the task, is it true, is it appropriate, is it mine? - Ask the AI to flag uncertain statements, open any sources it gives, and use tools for numbers. - Watch for stereotypes, follow privacy rules and workplace policies, and be open about AI use where required. - Build a personal prompt toolkit and keep updating it with prompts that work. ## Try it Create your prompt toolkit document with at least five prompts you will reuse, each with a note on when to use it and what to check. - [Previous: Ten everyday prompting mistakes (and fixes)](https://optimizeall.com/learn/prompt-engineering-foundations/everyday-mistakes) - [All lessons of Prompt Engineering Foundations](https://optimizeall.com/learn/prompt-engineering-foundations)