The parts of a prompt

A good prompt is not a longer prompt. It is one that leaves fewer decisions to be guessed at. There are six things a request can specify, and a weak prompt is usually one that omitted the two that mattered.

The six parts

Part	The question it answers	Example
Task	What should be done?	“Write a one-page summary”
Context	What should it be based on?	“…of the attached vendor contract”
Examples	What does good look like?	“Match the tone of last month’s memo”
Format	What shape should the answer take?	“Three bullets per section, no preamble”
Constraints	What must be true or avoided?	“Under 400 words, no legal advice”
Audience	Who is reading it?	“For the finance lead, who has not seen the contract”

Most requests need three or four of these, not all six.

From vague to specific

Weak: Summarise this.

Better: Summarise the attached contract for our finance lead in under 300 words. Lead with payment terms and termination. Flag anything unusual as a bullet at the end. Do not give legal advice.

The second one is not more polite or more elaborate. It has replaced four guesses — who, how long, what matters, what to avoid — with four answers.

Context is the part people skip

The single highest-return habit in Dhanu is attaching the actual material.

• Attach the file, or point at it with / in the composer
• Reference another task by its code, so its history comes with it
• Let the assistant search the Brain for your own policies instead of describing them from memory

An answer built on your documents is checkable. An answer built on a description of your documents is a guess with a citation-shaped hole in it.

Format is how you stop having to reformat

If you know the output will go into a table, a deck, or an email, say so up front. “As a table with columns: risk, likelihood, owner, mitigation” costs eight words and saves a round trip.

The same goes for what you do not want: “no preamble, no summary of what I asked” removes the two most common filler paragraphs.

Examples beat adjectives

“Professional” means different things to different teams. One paragraph of something you liked tells the model more than five adjectives, because it carries the vocabulary, the sentence length and the level of formality at once.

Two examples are much better than one — a single example gets copied, a pair gets generalised.

Constraints, and stating the negative

Say what must not happen when it matters: don’t invent figures, don’t email anyone, don’t change the existing sections, only use the attached sources. Models follow explicit prohibitions well, and infer them badly.

A template you can reuse

Do: <task>
Using: <the attached / linked material>
For: <audience>
As: <format>
Rules: <constraints>

Save the ones you keep retyping — see Saved prompts, which also support fill-in-the-blank fields so one prompt covers a whole category of request.

Next

Prompting techniques — the named moves that go beyond describing the task well.