Prompting
Writing better AI prompts for work | AI Course London
A useful work prompt supplies the source material, states the task, describes the output and sets boundaries. Role, context, task, format and constraints can serve as a practical checklist. They do not guarantee a correct answer. Start with a small result you can check, then revise the instruction when you can name a specific problem.
AI Course London · First published · Updated
What information belongs in the prompt?
Include the material the tool needs to perform the task. Name the reader, the decision or deliverable, and what a correct result must preserve. A professional-sounding role can help frame a draft, but it cannot supply missing expertise or factual evidence.
| Prompt part | Useful instruction | What it contributes |
|---|---|---|
| Role | Act as an editor of a customer update | A writing perspective |
| Context | The audience has already received the sample | The situation |
| Task | Draft an update from the supplied notes | The actual job |
| Format | At most 80 words, then missing facts | A reviewable shape |
| Constraints | Do not promise a delivery date | A boundary to check |
How does a vague prompt become a useful one?
Replace “make this professional” with an explicit source, audience, output and exclusion. The improved prompt should make it easier to tell whether the result is correct, not simply produce a longer answer.
Fictional supplier-update exercise
Source material
A supplier says the sample is ready. A courier has not been booked. Our customer asked for the tracking number.
Prompt to try
Using only these notes, draft a reply to our customer in at most 70 words. State that the sample is ready and that the courier booking is still outstanding. Do not invent a tracking number or delivery date. Then list what our team must confirm.
Human reference for checking
The reply says the sample is ready and the courier booking is outstanding. It does not claim dispatch, supply a tracking number or promise arrival. The team must confirm the courier booking and tracking details before sending an update with those facts.
This is a fictional practice exercise. The reference is written for checking your result; it is not an attendee outcome or a claimed model response.
What should I ask for when the source is incomplete?
Ask the tool to say what is missing and distinguish an assumption from a source fact. Request supporting phrases or source references for the important parts of the output. Ask for checks and unresolved questions rather than private internal reasoning.
If the task requires current information, identify how it will be obtained and verified. Do not let a request to “be confident” turn an unknown into a claim. A shorter answer that preserves a gap is more useful than an apparently complete invention.
How should I revise an answer that is almost right?
Name the defect precisely and preserve what already works. For example: “Keep the structure, remove the promised delivery date and add the courier booking to unresolved questions.” Ask for a replacement of the affected part rather than a complete rewrite when that is enough.
Check the revised answer against the original source, not just the previous draft. A revision can correct one problem and introduce another. Keep the version that passed review with its input and prompt so the example is reproducible.
Can I reuse a prompt across my team?
Reuse the task structure, but make the required inputs and checks explicit. Remove private examples from a shared template. Include a named review step and explain when a person should stop and ask for clarification.
Keep a short change log of failed cases and improved instructions. A prompt library is useful only when people understand its limits. Our workshop gives you practice adapting instructions to your own project; it does not promise a universal prompt that works every time.
Before you try it
Is role, context, task, format and constraints a proven universal formula?
It is a practical writing checklist used here to make instructions easier to review. Different tasks need different information, and the checklist does not establish a measurable success rate.
Should I ask the tool to show all its thinking?
Ask for relevant evidence, assumptions, checks and unresolved questions. These provide reviewable information without depending on private internal reasoning.
Can a better prompt remove the need to check the answer?
No. Check important facts and the output against the source. A well-structured request can make that checking easier, but it does not guarantee accuracy.
Check the source
Product features, plans and settings change. Check the provider and your organisation before using work material.
Provider information
Check the actual plan before choosing
Claude is the starting point in our public workshop. Check the current plan for your task.
Compare the product and plan with your required workflow and workplace permissions.
Check the current Notion AI features and plan terms if your team already uses Notion.
Claude and ChatGPT links go directly to the providers. The Notion link is an affiliate link; we may receive a commission if you make an eligible purchase.
Practise on your own work problem
Join Andrew for a one-day London workshop for non-technical professionals. Work in a group of at most eight, with guided practice from 9am to 4pm. The £599 fee includes lunch, materials and 30 days of support.
See workshop dates and details