Prompting on real work

Notes · slide 1

Welcome everyone. Say that this session is hands-on: by the end, everyone will have rewritten a prompt they actually use at work, and most people will see a noticeably better result in the room. Ask people to have your AI assistant open, and to think of something they asked it in the last week or two. If someone has never used it, that's fine: they can pick a task they'd like to try. Reassure the room there's no special "prompt engineering" vocabulary to learn. It's mostly about telling it what you'd tell a capable new colleague.

Notes · slide 2

30 seconds. Point out that the exercise takes about a third of the session, so the theory is short on purpose. If people haven't got your AI assistant open, ask them to open it now.

Notes · slide 3

This is the main idea. When people get a bland, generic answer, they often think they need cleverer wording or a magic phrase. Usually the real issue is that the AI has no idea who the output is for, what good looks like at your company, or what's already been decided. A capable new starter would have the same problem. Ask the room: "If you gave this task to a new colleague on their first day, what would you have to explain?" That's what belongs in the prompt. One minute.

Notes · slide 4

Read both aloud. If you can, run both live in your AI assistant and show the difference on screen; it's far more convincing than describing it. Point out what the second prompt adds: who it's for, what they care about, a length, a tone instruction, and an example of the style. None of that is clever wording, it's just information. Likely question: "Isn't that slower than writing the email myself?" Good answer: "The first time, maybe. But you can save this and reuse it, and once it works, filling in the details takes a minute." Two to three minutes.

Notes · slide 5

This is the framework to remember. Context is the biggest win. Constraints stop it rambling or overselling. An example is the fastest way to get your style: one good past email, report or reply says more than a paragraph of description. Make sure any example you paste is something you're allowed to share; we cover that in the data safety session. Ask the room for one prompt they used this week, and rewrite it live using these three headings. Pick a simple one. Three minutes.

Notes · slide 6

Give a function-specific example. HR: "This is a summary of the new hybrid working policy for line managers, who'll be asked questions by their teams. They need to know what changed, from when, and where to send exceptions." Support: "The customer has contacted us three times about the same delayed order and is angry. We can offer a refund but not expedited shipping." Notice that the support example includes what's already been decided, the refund but not shipping. Without that, AI will happily promise things your company can't deliver. Two minutes.

Notes · slide 7

Constraints are where you put your company's standards. If your company has a tone of voice guide, this is where it shows up. The "off-limits" line is the one people forget and the one that matters most for customer-facing work: for example "don't mention pricing", "don't admit liability", "don't use the word 'unfortunately'". Ask: "What does your team's 'don't' list look like?" Take one or two answers. Ninety seconds.

Notes · slide 8

One good example is worth more than a long description of your style. For a finance variance note, paste last month's commentary that the finance director liked, and say "match this structure and level of detail". For a sales email, paste one that got a reply. Important: tell it to match the style, not copy the content, or you'll get last month's facts back. Remove anything confidential from the example first. One to two minutes.

Notes · slide 9

Transition. Say: "The second big shift is not expecting the first answer to be the final one."

Notes · slide 10

Most people either accept the first answer or give up and start over. Both waste the work already done. Specific feedback works best: "the second paragraph overstates what we can offer, remove it" beats "make it better". "Ask me three questions before you start" is the single most useful habit for complex tasks: the questions often show you what context you forgot. Starting a fresh conversation is right when it's gone badly off track, but otherwise refine in place. Two minutes.

Notes · slide 11

Each example uses the same three ingredients: context, constraints, an example where you have one. The finance one is worth stressing: you provide the numbers, and AI helps explain and structure them. It should never be the source of the figures. If your audience is from one team, replace these with four tasks from that team; ask their manager beforehand. One to two minutes.

Notes · slide 12

Illustrative, not a real quotation. If you have a real, anonymised example from your company, use that. The lesson: "better" means nothing to the AI. Better how? Shorter? Warmer? More direct? Clearer call to action? Without being told, it tends to add words and polish. Ask the room: "How would you rewrite that instruction?" A good answer: "Cut this to under 100 words, keep the question at the end, make it sound like me talking." One minute.

Notes · slide 13

Timing: 12 minutes. Two minutes to find a prompt: people can scroll back through their your AI assistant history, or write down a task they'd like to do. Six minutes to rewrite it using the three headings and run both versions. Remind them not to paste anything confidential: use their judgement, or replace names with "Customer A". Walk around and help people who are stuck; the most common problem is people describing their task in general terms, so ask "who exactly is this for?" Four minutes for the next slide's pair discussion. If online, use breakout rooms of two or three. Have a visible timer. If someone finishes early, ask them to try "ask me three questions before you start" on the same task.

Notes · slide 14

4 minutes in pairs, then 2 to 3 minutes taking examples from the room. Ask "what was the most useful thing you added?" and note the answers; context nearly always wins. If someone's rewritten prompt didn't help, treat it as useful: ask what was still wrong, and get the room to suggest what context might be missing. Sometimes the honest answer is that the task is a poor fit, which links back to Session 1.

Notes · slide 15

This is the slide people photograph. Suggest they keep it next to them for a week until it becomes habit. Not every prompt needs every item; a quick rewording task needs very little. But anything going to a customer, a senior leader or a lot of people should tick most of the boxes. If your programme has a printable version or an intranet page, point to it here.

Notes · slide 16

The prompt you just rewrote is reusable. Turn the specifics into blanks, for example "[customer name]", "[their main concern]", "[paste figures]", and keep it somewhere you'll find it: a notes file, a team document, or a saved prompts feature if your AI assistant has one at your company. Check what's available before the session. When a whole team uses the same good prompt, that's the start of a shared library, which is how individual tricks become team capability. There's a quick lesson on turning a good prompt into a shared skill.

Notes · slide 17

One minute. Then take questions. Common question: "How long should a prompt be?" Good answer: "As long as it needs to be to give a new colleague enough to do the job. For a customer reply that might be a short paragraph plus the original message." Another: "Should I be polite to it?" Good answer: "Being clear matters far more than being polite. Write the way you'd brief a colleague."

Notes · slide 18

Make it specific: one prompt, one colleague, by Friday. Ask them to post their best rewritten prompt in the your AI programme channel or reply to your follow-up email, so you can start a shared collection. Set that up before the session. Thank everyone and mention the next session on what's safe to paste in.