Dragonfly Thinking · AI Foundations · Session 1

Don't write the prompt. Get one written.

A good prompt has up to six moving parts — role, task, context, examples, format, constraints. Building one from a blank page is slow, and most people stop at the task. This prompt does it for you: paste it, tell it what you want, answer three questions, and it writes the prompt — showing you which lever each part of it pulls.

How: copy it, paste it into ChatGPT, Claude, Copilot or Gemini, and send. Nothing to fill in first — it opens by asking what you want a prompt for, then asks the three things it most needs to know. Answer in a sentence each and it hands you the finished prompt.

Prompt · write me the prompt

You are a prompt engineer. Your job is to write the prompt I should send to an AI.

Start by asking me one question: what do I want a prompt for? Ask nothing else in that
first message — no options, no examples, no preamble.

When I answer, ask me the three things you most need to know to do this well — the
things I'd never think to tell you. Number them. Wait for my answers.

If I don't know an answer, or skip one, that's fine — take what I've given you and
carry on. Never send me away to find something first.

If I paste a prompt I already wrote, treat that as my answer and rewrite it.

Then write the prompt.

Build the prompt from these levers. Use only the ones that earn their place:

- Role — who the AI should be, when expertise or audience changes the answer
- Task — the one thing to do, stated as an instruction, not a topic
- Context — the background: my organisation, the project this belongs to, what it's
for in the wider scheme, who reads it, what's already been decided
- Examples — one short sample of the output, when the shape is hard to describe.
If I don't have one, that's fine — say so and move on. We can add one later.
- Format — the exact structure to return, headings and all
- Constraints — length, tone, what to leave out, what to do when something is missing

Rules for the prompt you write:
- Write it in my voice, addressed to the AI, ready to paste. No commentary inside it.
- Be specific where I was vague. If I said "summarise", say for whom and how long.
- Don't invent facts about my work. If a detail is missing and matters, leave a
{{PLACEHOLDER}} for me to fill in.
- Context and Examples are the two I'm most likely to be missing. If either would
change the answer for this task but I haven't given you the material, put the slot
in the prompt anyway, worded exactly like this, so I can fill it in later:
Context: {{BACKGROUND — my organisation, the project this is part of, what it's for}}
Example of what good looks like: {{PASTE ONE HERE, OR DELETE THIS LINE}}
Don't invent the content, and don't list it under "Left out" — a slot I can fill is
worth more than a note saying I didn't. If a lever wouldn't change the answer for
this task, leave it out entirely rather than adding an empty slot.
- Assume I'll paste my material underneath, so end the prompt with a line saying
where that goes.
- As short as it can be while still being unambiguous.

Return, in this order:
1. The prompt, in a code block.
2. **Levers used** — name each one and quote the words in the prompt that do its job.
Don't quote the same words for two levers; if a phrase does two jobs, name it once.
3. **Left out** — the levers that genuinely don't apply here, and why, one line each.
A lever I simply had no material for is a {{PLACEHOLDER}} above, not a line here.
4. **One upgrade** — the single change that would most improve the result.

Ask your three questions once. Don't come back for a second round — if something is
still missing, leave a {{PLACEHOLDER}} in the prompt instead.

Why three questions The questions are the point, not a delay. They pull out the context you would never think to volunteer — who reads this, what has already been decided, what happens if it is wrong — and that context is what separates a generic answer from a useful one. You don't need an answer to all three — say you don't know and it carries on; an example or a format can always be added later. It is a move you can use anywhere: before any hard question, try “before you answer, ask me the three things you most need to know.”

Read what comes back, don't just paste it. Along with the prompt you get three things worth more than the prompt itself:

Levers used — each part of your prompt, named and quoted. This is the anatomy card, applied to your own work rather than an example.

Left out — the levers it decided you didn't need, and why. Not every prompt needs all six; a short task with three levers beats a bloated one with six.

The slots it leaves — if background or an example would change the answer but you don't have one to hand, it leaves a labelled {{...}} in the prompt rather than dropping it or inventing it. Fill it in whenever you like; the prompt works without it.

One upgrade — the single change that would most improve the result. Make it, and run it again.