Capstone: use AI well for one real task

You've read about briefing, hallucinations, prompts, costs, and ethics. Now use them — once, deliberately, on something that matters to you this week.

Why this matters

Reading about prompting and doing prompting are different muscles. The track gave you the theory; this exercise builds the instinct. After one careful round, prompting stops feeling like incantation and starts feeling like briefing — which is what it actually is.

The challenge (~30 minutes)

Pick one real, slightly-ambitious task you've been putting off. Examples by role:

• Manager: write the agenda for next week's tense one-on-one.
• Engineer: explain a tricky piece of legacy code to a junior engineer joining next month.
• Marketer: turn a 2-page strategy doc into a 5-bullet exec summary.
• Student / researcher: outline the argument of a 30-page paper before reading it.
• Parent / volunteer / human in general: draft a difficult email you've been avoiding.

The task should be:

• Real — something you actually need to do, not contrived.
• Bounded — you can finish in 30 minutes.
• Worth doing well — has actual stakes (a person will read it, a decision rides on it).

The method

Step 1 — Pre-AI baseline (5 min)

Before you touch the chatbot, write down — in 3 bullets:

1. What does good look like for this task?
2. What's the format? (Email, doc, list, table, slide?)
3. What constraints matter? (Tone, length, things to avoid.)

This is your benchmark. If you skip this, you'll accept whatever the AI produces — even when "whatever" isn't actually good.

Step 2 — First prompt (5 min)

Use the briefing template from lesson 11:

[Role]      You're a __________________________________
[Task]      __________________________________________
[Style]     __________________________________________
[Format]    __________________________________________
[Input]     <your raw material here>

Send it. Read the output. Don't fix it yet.

Step 3 — Iterate (10 min)

Three short follow-ups, max:

• "More direct, drop the marketing speak."
• "This part is wrong because X. Try again with [correction]."
• "Now make it 30% shorter without losing the meaning."

Each round is one short sentence. Resist the urge to re-prompt from scratch.

Step 4 — Verify (5 min)

For anything that matters in the output:

• Numbers, dates, names: check against a source. Don't trust.
• Quotes / citations: search for them. Don't trust.
• Strong claims: ask yourself "would I bet £100 this is true?" If no, verify or hedge.
• Tone: does this sound like you? If a colleague received this, would it feel right?

Step 5 — Reflect (5 min)

Write 3 bullets:

1. What did AI genuinely help with? (Be specific.)
2. Where did you still have to do the work? (The judgement, the verification, the final polish — usually.)
3. What would you do differently next time? (One concrete change.)

Self-check

Answer these honestly:

What's next

You've completed the foundations. From here, two doors:

• Stay non-technical: keep practising this method on real tasks. Read Choosing the right AI tool and Privacy again as needs arise. Re-read Hallucinations when something feels off.
• Go deeper: the Medium track covers how production AI products are actually built — embeddings, RAG, agents, evals. Beginner concepts you've met will come back, in code form.

Either path is legitimate. The point isn't to finish the curriculum — it's to use AI well.